Jev AtlasTypeSafe & System One
About the directory GitHub
THE COMMUNITY RESOURCE DIRECTORY

Small decisions.
Endless possibilities.

Find what people build with Jev. Explore tools, libraries, experiments and ideas from across the TypeSafe ecosystem.

Updated · 2026-10-01About the directory
11,066unique resources
29clear categories
220focused subcategories
Yourssave a personal reading list
YOUR NEXT BUILD STARTS HERE

Explore the collection

Download catalog

569 of 11,066 resources

Open models & compatible servers569 of 569 matches
Open implementations137 of 137 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
AgentJevhuggingface.co—Open 0.6B agent decision model with shared-prefix caching that returns distributions for questions like which tool is next; 79.25% top-1 on 2,000 Typed Decisions questions vs 72.7% for Jev 1.13.0.Open models & compatible serversOpen implementations
ai_sdkgithub.com2,462TianyuCodings/NanoJev — A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline. · ai_sdk · PythonOpen models & compatible serversOpen implementations
biviogithub.com0Decisioni tipizzate da un modello linguistico, sul tuo computer, senza generare un token. La versione libera dell'idea di Jev.Open models & compatible serversOpen implementations
Build a Jev from scratchhuggingface.co—Toy 3.1M-parameter reconstruction of the System One interface with training, eval and serving code plus an article; accuracy is low by design but calibration reaches ECE 0.027.Open models & compatible serversOpen implementations
Cerebellum-2Bgithub.com1Open-weight 2B decision model built on Qwen3.5-2B that picks tool routes and DOM actions from a candidate set in a single forward pass, pitched as an open alternative to Jev, with BF16, FP8 and INT8 weights.Open models & compatible serversOpen implementations
Cerebellum-2B — huggingfacehuggingface.co—Open-weight 2B decision model built on Qwen3.5-2B that picks tool routes and DOM actions from a candidate set in a single forward pass, pitched as an open alternative to Jev, with BF16, FP8 and INT8 weights.Open models & compatible serversOpen implementations
ChristianAlexander/effect-jev-cwegithub.com1A demonstration of the Jev System 1 model in Effect, matching vulnerabilities to their underlying CWEsOpen models & compatible serversOpen implementations
CLMgithub.com2,658Open-source contrastive System One decision model with a TypeSafe-compatible API for typed Noul, Choice, and Score judgments.Open models & compatible serversOpen implementations
decidergithub.com1,023Open System One-style models fine-tuned from Qwen3.5, a 2B and a 35B mixture-of-experts, that return a probability distribution for every typed question in one forward pass, demoed on ten text games and Super Mario Bros.Open models & compatible serversOpen implementations
decider — modelhuggingface.co—Open System One-style models fine-tuned from Qwen3.5, a 2B and a 35B mixture-of-experts, that return a probability distribution for every typed question in one forward pass, demoed on ten text games and Super Mario Bros.Open models & compatible serversOpen implementations
decimagithub.com0Open, CPU-sized Jev-style (System One) decision model: situation + question + options → calibrated probabilities. 122M, ~20 ms on one CPU core, runs in your browser.Open models & compatible serversOpen implementations
DeepOpengithub.com59Laya-based open System 1 decision engine with a Chinese technical whitepaper, a script/language router over English, multilingual and typed-decision checkpoints, and training experiments on CLINC150 and Banking77.Open models & compatible serversOpen implementations
DeepOpen — appdeepopen.com—Laya-based open System 1 decision engine with a Chinese technical whitepaper, a script/language router over English, multilingual and typed-decision checkpoints, and training experiments on CLINC150 and Banking77.Open models & compatible serversOpen implementations
djev devgithub.com93Open implementation of typed decisions on DiffusionGemma and vLLM that reads label probabilities from denoised answer positions, with image inputs, image choices and live camera sampling.Open models & compatible serversOpen implementations
Dohnutsgithub.com18Small multimodal System One style models, with a 0.8B checkpoint that runs on a consumer GPU, answering choice, noul and score questions about text or images in one forward pass, plus training and eval toolkit.Open models & compatible serversOpen implementations
Dohnuts — sitedohnuts.ai—Small multimodal System One style models, with a 0.8B checkpoint that runs on a consumer GPU, answering choice, noul and score questions about text or images in one forward pass, plus training and eval toolkit.Open models & compatible serversOpen implementations
endpointgithub.com160kshetrajna12/reflex — A small open decision model: state + typed questions -\> calibrated probabilities. A Jev / System One re-creation on Qwen3.5. · endpoint · PythonOpen models & compatible serversOpen implementations
endpointgithub.com797wfzyx/von — The open-source System One decision model. Sub-15ms, non-autoregressive, local drop-in alternative to TypeSafe Jev. · endpoint · PythonOpen models & compatible serversOpen implementations
Foqgithub.com5Local, open-source Jev alternative that returns calibrated probabilities on typed boolean, choice, score and Pydantic questions in one feed-forward pass, reported at about 25 ms, with benchmarks against Laya.Open models & compatible serversOpen implementations
Foq — appfoq.fr—Local, open-source Jev alternative that returns calibrated probabilities on typed boolean, choice, score and Pydantic questions in one feed-forward pass, reported at about 25 ms, with benchmarks against Laya.Open models & compatible serversOpen implementations
imajevgithub.com186Open Jev-style typed-decision model that also takes images: photo + app state + typed questions in, calibrated probabilities out, locally.Open models & compatible serversOpen implementations
Jev-Visionhuggingface.co—Open-weight vision decision model (LoRA plus typed heads on Qwen3-VL-8B) that answers typed questions about screenshots and images through a Jev-compatible /v1/systemone API, 150-190 ms per screen step.Open models & compatible serversOpen implementations
Jev-Vision — projectgithub.com4Open-weight vision decision model (LoRA plus typed heads on Qwen3-VL-8B) that answers typed questions about screenshots and images through a Jev-compatible /v1/systemone API, 150-190 ms per screen step.Open models & compatible serversOpen implementations
JevBERT (typed-decision-bert)github.com20Unofficial proof of concept of a BERT-style encoder decision engine behind a Jev-shaped /v1/systemone API for noul, choice, and score questions, with published results and a benchmark against real Jev.Open models & compatible serversOpen implementations
JevBERT (typed-decision-bert) — demox.com—Unofficial proof of concept of a BERT-style encoder decision engine behind a Jev-shaped /v1/systemone API for noul, choice, and score questions, with published results and a benchmark against real Jev.Open models & compatible serversOpen implementations
JevK5github.com128Qwen3.5 replica with distilled LoRA weights, ranked second of 76 systems and first among open ones on JevBench v1.4.Open models & compatible serversOpen implementations
Jevlikegithub.com1,338Independent starter model with the same shape as Jev: a small one-pass scorer mapping text plus N options to one probability per option, with Doom and chess vision demos and a Wikispeedia next-click example.Open models & compatible serversOpen implementations
Jevlike — demox.com—Independent starter model with the same shape as Jev: a small one-pass scorer mapping text plus N options to one probability per option, with Doom and chess vision demos and a Wikispeedia next-click example.Open models & compatible serversOpen implementations
Jevlike — discussionnews.ycombinator.com—Independent starter model with the same shape as Jev: a small one-pass scorer mapping text plus N options to one probability per option, with Doom and chess vision demos and a Wikispeedia next-click example.Open models & compatible serversOpen implementations
Jevonhuggingface.co—Tiny 20M-parameter decision model that answers typed questions about a grid (which way to move, is north clear, how far is the goal) for maze and snake, rebuilt from NanoJev around planning depth.Open models & compatible serversOpen implementations
Jevon — projectgithub.com—Tiny 20M-parameter decision model that answers typed questions about a grid (which way to move, is north clear, how far is the goal) for maze and snake, rebuilt from NanoJev around planning depth.Open models & compatible serversOpen implementations
jevosgithub.com1,154Open-source, CPU-only alternative to Jev for yes/no decisions: MiniCPM5-1B cut to 17 layers with a one-logit output head, quantized to a 619 MB GGUF, one forward pass on llama.cpp, no GPU.Open models & compatible serversOpen implementations
jevos — releasegithub.com1,154Open-source, MIT-licensed alternative to Jev for yes/no questions, running fully offline on a laptop CPU in one forward pass with no text generation.Open models & compatible serversOpen implementations
jqvgithub.com—Reconstruction of Jev's inference structure on stock Qwen3, with one shared state prefill, isolated question branches behind a block attention mask, and option-letter logits calibrated by temperature, behind /v1/systemone.Open models & compatible serversOpen implementations
Kev — demox.com—Open 0.8B to 9B decision models on Qwen3.5 with released weights and a server compatible with /v1/systemone, so the official SDKs can point at it.Open models & compatible serversOpen implementations
Kev — discussionnews.ycombinator.com—Open 0.8B to 9B decision models on Qwen3.5 with released weights and a server compatible with /v1/systemone, so the official SDKs can point at it.Open models & compatible serversOpen implementations
Kev — modelhuggingface.co—Open 0.8B to 9B decision models on Qwen3.5 with released weights and a server compatible with /v1/systemone, so the official SDKs can point at it.Open models & compatible serversOpen implementations
Kev — model 2huggingface.co—Open 0.8B to 9B decision models on Qwen3.5 with released weights and a server compatible with /v1/systemone, so the official SDKs can point at it.Open models & compatible serversOpen implementations
Layagithub.com12,037Open, local non-autoregressive decision models on ModernBERT-style encoders answering Choice, Score and Noul in about 33 ms per question; base weights score below majority zero-shot and ship over-confident, so fine-tune first.Open models & compatible serversOpen implementations
Laya Visiongithub.com21Research fork of the open Laya model that swaps its text encoder for SmolVLM-256M, so typed choice, score, and noul decisions can be read from an image plus optional text in one forward pass.Open models & compatible serversOpen implementations
Laya Visionwww.reddit.com—Open Jev-style decision model for images built on SmolVLM 256M, returning probabilities over a fixed set of choices for screenshots, photos, and GUI elements.Open models & compatible serversOpen implementations
Laya Vision — apphuggingface.co—Research fork of the open Laya model that swaps its text encoder for SmolVLM-256M, so typed choice, score, and noul decisions can be read from an image plus optional text in one forward pass.Open models & compatible serversOpen implementations
Laya Vision — discussionnews.ycombinator.com—Research fork of the open Laya model that swaps its text encoder for SmolVLM-256M, so typed choice, score, and noul decisions can be read from an image plus optional text in one forward pass.Open models & compatible serversOpen implementations
Laya Vision — modelhuggingface.co—Research fork of the open Laya model that swaps its text encoder for SmolVLM-256M, so typed choice, score, and noul decisions can be read from an image plus optional text in one forward pass.Open models & compatible serversOpen implementations
Laya — apphuggingface.co—Open, local non-autoregressive decision models on ModernBERT-style encoders answering Choice, Score and Noul in about 33 ms per question; base weights score below majority zero-shot and ship over-confident, so fine-tune first.Open models & compatible serversOpen implementations
Laya — articledev.to—Open, local non-autoregressive decision models on ModernBERT-style encoders answering Choice, Score and Noul in about 33 ms per question; base weights score below majority zero-shot and ship over-confident, so fine-tune first.Open models & compatible serversOpen implementations
Laya — discussionwww.reddit.com—Open, local non-autoregressive decision models on ModernBERT-style encoders answering Choice, Score and Noul in about 33 ms per question; base weights score below majority zero-shot and ship over-confident, so fine-tune first.Open models & compatible serversOpen implementations
Laya — discussion 2www.reddit.com—Open, local non-autoregressive decision models on ModernBERT-style encoders answering Choice, Score and Noul in about 33 ms per question; base weights score below majority zero-shot and ship over-confident, so fine-tune first.Open models & compatible serversOpen implementations
Laya — modelhuggingface.co—Open, local non-autoregressive decision models on ModernBERT-style encoders answering Choice, Score and Noul in about 33 ms per question; base weights score below majority zero-shot and ship over-confident, so fine-tune first.Open models & compatible serversOpen implementations
Laya-CoreMLgithub.com989Open-weight port of the Jev-like Laya model to Apple Core ML and the Neural Engine, serving the System One contract locally, with a Snake demo playing on live probabilities.Open models & compatible serversOpen implementations
Laya-CoreML — pypipypi.org—Open-weight port of the Jev-like Laya model to Apple Core ML and the Neural Engine, serving the System One contract locally, with a Snake demo playing on live probabilities.Open models & compatible serversOpen implementations
Laya-CoreML — repohuggingface.co—Open-weight port of the Jev-like Laya model to Apple Core ML and the Neural Engine, serving the System One contract locally, with a Snake demo playing on live probabilities.Open models & compatible serversOpen implementations
LCTgithub.com5Jev-LCT: Open System-One Decision Engine with Free Calibrated Confidence from Recurrent TrajectoriesOpen models & compatible serversOpen implementations
leCore holographic System Onegithub.com—Native implementation of Jev's state-plus-typed-questions contract inside a pure-NumPy vector engine, as a few-shot prototype classifier that abstains on ties and returns no probability until calibrated.Open models & compatible serversOpen implementations
leCore holographic System One — appdiscoverleos.com—Native implementation of Jev's state-plus-typed-questions contract inside a pure-NumPy vector engine, as a few-shot prototype classifier that abstains on ties and returns no probability until calibrated.Open models & compatible serversOpen implementations
leCore holographic System One — repogithub.com—Native implementation of Jev's state-plus-typed-questions contract inside a pure-NumPy vector engine, as a few-shot prototype classifier that abstains on ties and returns no probability until calibrated.Open models & compatible serversOpen implementations
levgithub.com43An open System One decision modelOpen models & compatible serversOpen implementations
llama-cpp-system-onegithub.com0A Rust implementation of the System One API for structured question answering with DiffusionGemma and llama.cpp (OpenJev - Open, Jev Compatible).Open models & compatible serversOpen implementations
Local Jev experimentx.com—Demo of a homemade Jev-style decision setup running on a local LLM, with the author noting its speed still has room to improve.Open models & compatible serversOpen implementations
malevrigns/agent-jevgithub.com333AgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any unstructured state (diffs, traces, logs) and structured questions, get calibrated probability distributions back in one ~50ms forward pass. Zero output-token decoding.Open models & compatible serversOpen implementations
metask-jevgithub.com2Open-weight Jev-class typed-decision models on Qwen3.5 that read calibrated per-option probabilities in one forward pass; metask-jev-4b reports 80.1% on JevBench versus 75.3% for Jev 1.13.0.Open models & compatible serversOpen implementations
metask-jev — modelhuggingface.co—Open-weight Jev-class typed-decision models on Qwen3.5 that read calibrated per-option probabilities in one forward pass; metask-jev-4b reports 80.1% on JevBench versus 75.3% for Jev 1.13.0.Open models & compatible serversOpen implementations
metask-jev — model 2huggingface.co—Open-weight Jev-class typed-decision models on Qwen3.5 that read calibrated per-option probabilities in one forward pass; metask-jev-4b reports 80.1% on JevBench versus 75.3% for Jev 1.13.0.Open models & compatible serversOpen implementations
model_idgithub.com1DECRUX9812/openjev-lm — open-Jev LM arm: Qwen2.5-0.5B + LoRA reproducing a hosted decision model's judgment at 92.9% on hand-labelled gold - trained overnight on a 6-vCPU CPU-only host, $0/call. Paper, corpora, harnesses, receipts. · model_id · PythonOpen models & compatible serversOpen implementations
model_idgithub.com109Heman10x-NGU/Verdict-open-jev — Non-autoregressive decision engine on ModernBERT (151M) with calibrated uncertainty (RLCD), TypeSafe AI Jev benchmark audit, and in-browser WebGPU playground · model_id · PythonOpen models & compatible serversOpen implementations
model_idgithub.com293Heman10x-NGU/openJev-verdict-2.0 — Calibrated 151M Non-Autoregressive Decision Engine beating TypeSafe Jev & Laya on LocalLLaMA/typed-decisions (77.10% acc, 0.0636 Brier, 0.0144 ECE) · model_id · PythonOpen models & compatible serversOpen implementations
model_idgithub.com6akash-kamat/system-one-gemma — Open-source Jev-style System One decision model. Gemma 3 270M with a scoring head — fast, calibrated decisions in a single forward pass. No text generation. Inspired by TypeSafe.ai's Jev. · model_id · PythonOpen models & compatible serversOpen implementations
model_idgithub.com58ikermoel/open-alternative-jev — Open-source alternative to TypeSafe's Jev: a System One style model layer that gives typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with honest benchmarks · model_id · PythonOpen models & compatible serversOpen implementations
model_idgithub.com24sabeel111/OpenSourceJev — Turning an LLM model into a Jev like System. · model_id · PythonOpen models & compatible serversOpen implementations
NanoJevgithub.com2,462Open 0.6B Jev replica trained for four specific games, with weights and data; not a general-purpose substitute.Open models & compatible serversOpen implementations
NanoJev — demox.com—Open 0.6B Jev replica trained for four specific games, with weights and data; not a general-purpose substitute.Open models & compatible serversOpen implementations
NanoJev — modelhuggingface.co—Open 0.6B Jev replica trained for four specific games, with weights and data; not a general-purpose substitute.Open models & compatible serversOpen implementations
Open Jevgithub.com61Unofficial PyTorch reconstruction of the ideas behind Jev with random weights: the state is encoded once and every question is answered through small typed readout heads in one parallel forward pass.Open models & compatible serversOpen implementations
Open Spark Jevgithub.com21Open local decision model, spark-s1, inspired by Jev and built on Qwen3 for NVIDIA DGX Spark, with a Jev-compatible API, architecture notes, data generation and evaluation tooling.Open models & compatible serversOpen implementations
Open vision Jev alternativesx.com—Tests open-source Jev-like models that accept images, finding DiffusionGemma (via a vLLM PR) and reflex on the Pareto front for vision decisions.Open models & compatible serversOpen implementations
Open-Jevgithub.com381Open probability-decision models on Qwen3.5-2B, Qwen3.5-9B, and Qwen3.8-27B that score supplied candidates without generation, released as LoRA adapters with a trained decision head, a dataset, and benchmarks.Open models & compatible serversOpen implementations
Open-Jev — demox.com—Open probability-decision models on Qwen3.5-2B, Qwen3.5-9B, and Qwen3.8-27B that score supplied candidates without generation, released as LoRA adapters with a trained decision head, a dataset, and benchmarks.Open models & compatible serversOpen implementations
Open-Jev — modelhuggingface.co—Open probability-decision models on Qwen3.5-2B, Qwen3.5-9B, and Qwen3.8-27B that score supplied candidates without generation, released as LoRA adapters with a trained decision head, a dataset, and benchmarks.Open models & compatible serversOpen implementations
Open-Jev — model 2huggingface.co—Open probability-decision models on Qwen3.5-2B, Qwen3.5-9B, and Qwen3.8-27B that score supplied candidates without generation, released as LoRA adapters with a trained decision head, a dataset, and benchmarks.Open models & compatible serversOpen implementations
Open-Jev — sitezefan-cai.github.io—Open probability-decision models on Qwen3.5-2B, Qwen3.5-9B, and Qwen3.8-27B that score supplied candidates without generation, released as LoRA adapters with a trained decision head, a dataset, and benchmarks.Open models & compatible serversOpen implementations
open-jev-deberta-v3-largehuggingface.co—Open Jev-shaped model on DeBERTa-v3-large that answers any number of choice, score and noul questions about one state in a single pass; 0.854 in-domain and 0.690 out-of-domain accuracy, 28 ms for 10 questions.Open models & compatible serversOpen implementations
open-jev-deberta-v3-large — apphuggingface.co—Open Jev-shaped model on DeBERTa-v3-large that answers any number of choice, score and noul questions about one state in a single pass; 0.854 in-domain and 0.690 out-of-domain accuracy, 28 ms for 10 questions.Open models & compatible serversOpen implementations
open-jev-deberta-v3-large — modelhuggingface.co—Open Jev-shaped model on DeBERTa-v3-large that answers any number of choice, score and noul questions about one state in a single pass; 0.854 in-domain and 0.690 out-of-domain accuracy, 28 ms for 10 questions.Open models & compatible serversOpen implementations
OpenDecisiongithub.com57Open-source local semantic decision engine modeled on Jev that answers Choice, Noul and Score questions plus a supports/contradicts Relation over structured state and documents, demoed playing Doom.Open models & compatible serversOpen implementations
OpenDecision — linkdeepanwadhwa.github.io—Open-source local semantic decision engine modeled on Jev that answers Choice, Noul and Score questions plus a supports/contradicts Relation over structured state and documents, demoed playing Doom.Open models & compatible serversOpen implementations
OpenJevx.com—Open-source on-device inference for Jev-style typed decisions with probabilities, built on Qwen 3 4B, 0.6B and 0.8B, and able to run in Chrome.Open models & compatible serversOpen implementations
OpenJev (SiliconLabAI)github.com67Open System One-style decision playground that answers a state and typed questions through LLM micro-scorers, one structured JSON call to an OpenAI-compatible model, or the open Mapika/decider weights.Open models & compatible serversOpen implementations
OpenJev (SiliconLabAI) — videowww.youtube.com—Open System One-style decision playground that answers a state and typed questions through LLM micro-scorers, one structured JSON call to an OpenAI-compatible model, or the open Mapika/decider weights.Open models & compatible serversOpen implementations
OpenJev (Verdict)github.com109Open 151M-parameter ModernBERT decision engine inspired by Jev and RLCD that evaluates typed questions in one non-autoregressive pass in under 35 ms, with a JevBench audit and an in-browser WebGPU playground.Open models & compatible serversOpen implementations
OpenJev (Verdict) — huggingfacehuggingface.co—Open 151M-parameter ModernBERT decision engine inspired by Jev and RLCD that evaluates typed questions in one non-autoregressive pass in under 35 ms, with a JevBench audit and an in-browser WebGPU playground.Open models & compatible serversOpen implementations
OpenJev — repogithub.com1,307Open-source on-device inference for Jev-style typed decisions with probabilities, built on Qwen 3 4B, 0.6B and 0.8B, and able to run in Chrome.Open models & compatible serversOpen implementations
openJev-verdict-2.0github.com293ModernBERT-based 151M non-autoregressive decision model with separate distribution and confidence heads, public weights, saved evaluation artifacts and an in-browser WebGPU demo; its benchmark results are self-reported.Open models & compatible serversOpen implementations
openJev-verdict-2.0 — discussionwww.reddit.com—ModernBERT-based 151M non-autoregressive decision model with separate distribution and confidence heads, public weights, saved evaluation artifacts and an in-browser WebGPU demo; its benchmark results are self-reported.Open models & compatible serversOpen implementations
openJev-verdict-2.0 — docsheman10x-ngu.github.io—ModernBERT-based 151M non-autoregressive decision model with separate distribution and confidence heads, public weights, saved evaluation artifacts and an in-browser WebGPU demo; its benchmark results are self-reported.Open models & compatible serversOpen implementations
OpenJev-Visiongithub.com28Open research toolkit for visual probabilistic decisions that encodes an image once and answers several structured questions from shared probabilities, with synthetic, Oxford-IIIT Pet, and CLEVR data and trained weights.Open models & compatible serversOpen implementations
OpenSourceJevgithub.com24Research experiment that runs a local System One-style decision engine on llama.cpp and Qwen, projecting next-token logits onto answer options for calibrated typed decisions on consumer hardware.Open models & compatible serversOpen implementations
OpenSourceJev — discussionwww.reddit.com—Research experiment that runs a local System One-style decision engine on llama.cpp and Qwen, projecting next-token logits onto answer options for calibrated typed decisions on consumer hardware.Open models & compatible serversOpen implementations
OpenThai-SystemOnegithub.com65Open Thai and English System One model built from a Qwen3.5-0.8B tower with a 256-way slot head, answering choice, score, and noul questions in one pass behind TypeSafe's /v1/systemone contract.Open models & compatible serversOpen implementations
OpenThai-SystemOne — apphuggingface.co—Open Thai and English System One model built from a Qwen3.5-0.8B tower with a 256-way slot head, answering choice, score, and noul questions in one pass behind TypeSafe's /v1/systemone contract.Open models & compatible serversOpen implementations
OpenThai-SystemOne — discussionwww.reddit.com—Open Thai and English System One model built from a Qwen3.5-0.8B tower with a 256-way slot head, answering choice, score, and noul questions in one pass behind TypeSafe's /v1/systemone contract.Open models & compatible serversOpen implementations
OpenThai-SystemOne — modelhuggingface.co—Open Thai and English System One model built from a Qwen3.5-0.8B tower with a 256-way slot head, answering choice, score, and noul questions in one pass behind TypeSafe's /v1/systemone contract.Open models & compatible serversOpen implementations
primitivesgithub.com38mithalouni/system-one-open — Open replica of TypeSafe's Jev: typed calibrated decisions in one forward pass, on Gemma 4 E2B / Gemma 3 270M (Modal) · primitives · PythonOpen models & compatible serversOpen implementations
primitivesgithub.com5xingwudao/OpenJev — OpenJev: an independent Jev-inspired System One decision API based on TypeSafe.ai concepts. Choice, score and noul primitives, local mock server, Python and TypeScript SDKs. Real inference planned; not affiliated with TypeSafe AI. · primitives · PythonOpen models & compatible serversOpen implementations
primitivesgithub.com5yohanargentina-oss/Foq — ⚡ Foq — the FREE, local, open-source alternative to Jev. Typed System 1 decisions in ~25 ms — no waitlist, no cloud, no per-token cost. foq.fr · primitives · PythonOpen models & compatible serversOpen implementations
reflexgithub.com160Open Jev re-creation on stock Qwen3.5 that answers Choice, Score and Noul questions in one batched pass behind a /v1/systemone server, with a WebGPU demo and a LoRA recipe for honest probabilities; ~200 ms per request.Open models & compatible serversOpen implementations
reflex — appkshetrajna12.github.io—Open Jev re-creation on stock Qwen3.5 that answers Choice, Score and Noul questions in one batched pass behind a /v1/systemone server, with a WebGPU demo and a LoRA recipe for honest probabilities; ~200 ms per request.Open models & compatible serversOpen implementations
rlx-kevgithub.com—Rust port of Kev, a small Jev-like System One decision model with a Qwen3.5 backbone, merged LoRA and pointer head, running natively on the RLX multi-backend ML compiler and runtime.Open models & compatible serversOpen implementations
rlx-kev — repogithub.com—Rust port of Kev, a small Jev-like System One decision model with a Qwen3.5 backbone, merged LoRA and pointer head, running natively on the RLX multi-backend ML compiler and runtime.Open models & compatible serversOpen implementations
rulinggithub.com6Open-weight recreation of Jev's typed-decision interface that runs on an Apple Silicon Mac with MLX and a 4-bit Qwen3.5 checkpoint, answering in about 119 ms; a 19 GB MoE model reaches Jev's accuracy on TypeSafe's evaluation rows.Open models & compatible serversOpen implementations
SemIf (formerly OpenJev) — demox.com—Independent Jev-like project that answers semantic if-questions by reading option logits from a frozen 4B open model on a home RTX 3090, with a browser demo and no text generation.Open models & compatible serversOpen implementations
SemIf (formerly OpenJev) — discussionwww.reddit.com—Independent Jev-like project that answers semantic if-questions by reading option logits from a frozen 4B open model on a home RTX 3090, with a browser demo and no text generation.Open models & compatible serversOpen implementations
Six Jev clones in two daysx.com—Japanese summary comparing the architectures of six open Jev clones: Laya, DiffusionGemmaJev, Bespoke Nimble, SemIf (formerly OpenJev), Jevlike and Kev-0.5B.Open models & compatible serversOpen implementations
Six Jev clones in two days — articlewww.latent.space—Japanese summary comparing the architectures of six open Jev clones: Laya, DiffusionGemmaJev, Bespoke Nimble, SemIf (formerly OpenJev), Jevlike and Kev-0.5B.Open models & compatible serversOpen implementations
smalljevgithub.com—Open Jev-style decision model built on MiniCPM5-2B-Base with a LoRA adapter and native decision heads, about 2.5B parameters answering in one forward pass with no generated tokens, aimed at phones.Open models & compatible serversOpen implementations
smalljev — modelhuggingface.co—Open Jev-style decision model built on MiniCPM5-2B-Base with a LoRA adapter and native decision heads, about 2.5B parameters answering in one forward pass with no generated tokens, aimed at phones.Open models & compatible serversOpen implementations
syngithub.com0Open-source, Jev-compatible System One API: typed decisions (Noul, Choice, Score) from next-token probabilities in one forward pass. Works with typesafe-sdk. Free hosted at sifty.dev.Open models & compatible serversOpen implementations
System One (phase2 student)huggingface.co—Local decision model from ModernBERT-base with LoRA and a two-layer scoring head that takes JSON or text state and answers typed choice, score and noul questions in one pass.Open models & compatible serversOpen implementations
System One (phase2 student) — projectgithub.com—Local decision model from ModernBERT-base with LoRA and a two-layer scoring head that takes JSON or text state and answers typed choice, score and noul questions in one pass.Open models & compatible serversOpen implementations
System One Gemmagithub.com6Open Jev-style decision model that puts a linear scoring head on Gemma 3 270M, scoring every option's sequence in one batched forward pass and softmaxing into calibrated Choice, Noul and Score answers.Open models & compatible serversOpen implementations
system-one-270mhuggingface.co—Open System One reproduction on Gemma 3 270M that takes state plus a typed question with caller-supplied options and returns one option with a calibrated probability, never free text.Open models & compatible serversOpen implementations
system-one-270m — projectgithub.com—Open System One reproduction on Gemma 3 270M that takes state plus a typed question with caller-supplied options and returns one option with a calibrated probability, never free text.Open models & compatible serversOpen implementations
system-one-opengithub.com38Open Jev-style decision model built on Gemma 4 E2B (attention LoRA) and Gemma 3 270M, trained and served on Modal; reports 76.7% on TypeSafe's public eval strict subset versus 86.9% for Jev.Open models & compatible serversOpen implementations
systemANEgithub.com8Proof of concept of a System One-style engine on Apple's Neural Engine with Choice, Boolean, and Score primitives over a 22.6M-parameter encoder, escalating unsure cases to Apple's Foundation Models.Open models & compatible serversOpen implementations
systemone-litegithub.com3Toy local approximation of the System One API backed by a small causal LM (Qwen2.5-0.5B by default), scoring only option tokens over a shared KV-cached state, with optional fine-tuned checkpoints.Open models & compatible serversOpen implementations
systemone-lite — demox.com—Toy local approximation of the System One API backed by a small causal LM (Qwen2.5-0.5B by default), scoring only option tokens over a shared KV-cached state, with optional fine-tuned checkpoints.Open models & compatible serversOpen implementations
systemone-lite — modelhuggingface.co—Toy local approximation of the System One API backed by a small causal LM (Qwen2.5-0.5B by default), scoring only option tokens over a shared KV-cached state, with optional fine-tuned checkpoints.Open models & compatible serversOpen implementations
systemone-lite — model 2huggingface.co—Toy local approximation of the System One API backed by a small causal LM (Qwen2.5-0.5B by default), scoring only option tokens over a shared KV-cached state, with optional fine-tuned checkpoints.Open models & compatible serversOpen implementations
tamevgithub.com0Open-source multi-device Jev-like decision models you can train and run yourself: 5.33 ms on one CPU thread. Apache-2.0.Open models & compatible serversOpen implementations
Tiny-Jevhuggingface.co—Open 0.6B System One model on Qwen3-0.6B that answers typed questions with calibrated probabilities: 95.8% accuracy and 0.004 ECE on in-family tests, 93.5% out of distribution.Open models & compatible serversOpen implementations
truetype.ai-opengithub.com3Open source replica of Truetype AI using open weight modelsOpen models & compatible serversOpen implementations
typedecidegithub.com0Open-source implementation of the typed-decision pattern popularised by TypeSafe's Jev: read a decision out of a small language model's logits, in the browser. Library, benchmarks and paper. Not affiliated with TypeSafe.Open models & compatible serversOpen implementations
Visual Jevgithub.com2Open, local image-native typed decisions on a Qwen3-VL backbone, answering Choice, Score and Noul questions about images as calibrated distributions without generating text.Open models & compatible serversOpen implementations
Visual-Jevgithub.com7Copy of the SemIf (formerly OpenJev) open reimplementation of Jev's typed-decision interface that adds an optional image input, running Qwen3.5-4B as a vision-language model for direct, serial and shared scoring.Open models & compatible serversOpen implementations
Visual-Jev — upstreamgithub.com4,638Copy of the SemIf (formerly OpenJev) open reimplementation of Jev's typed-decision interface that adds an optional image input, running Qwen3.5-4B as a vision-language model for direct, serial and shared scoring.Open models & compatible serversOpen implementations
Vongithub.com797Open non-autoregressive decision model of about 395M parameters with public weights, training code and a System One-shaped API, answering Choice, Noul and Score questions locally in tens of milliseconds.Open models & compatible serversOpen implementations
Von — discussionnews.ycombinator.com—Open non-autoregressive decision model of about 395M parameters with public weights, training code and a System One-shaped API, answering Choice, Noul and Score questions locally in tens of milliseconds.Open models & compatible serversOpen implementations
xingwudao/OpenJevgithub.com5OpenJev: an independent Jev-inspired System One decision API based on TypeSafe.ai concepts. Choice, score and noul primitives, local mock server, Python and TypeScript SDKs. Real inference planned; not affiliated with TypeSafe AI.Open models & compatible serversOpen implementations
Training & distillation29 of 29 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
decision-head-rlcdgithub.com6Experiment training Qwen3.5-4B with a rank-8 LoRA and RLCD on 32,000 typed decisions to test where a Jev-like model's generalisation comes from; it finds generalisation tracks training-data coverage.Open models & compatible serversTraining & distillation
Distilled 4B local decision modelx.com—Distillation of DeepSeek V4 Flash judgments into a Jev-style 4B local model over 26 hours on a DGX Spark, beating the teacher's instant mode at 1/20 the size with about 22ms per decision.Open models & compatible serversTraining & distillation
eve-rlcdgithub.com4Jev-inspired 0.6B decision model on Qwen3-0.6B-Base trained with RL from right/wrong feedback so its option probabilities match accuracy; includes training loop, ablation, eval and open weights.Open models & compatible serversTraining & distillation
eve-rlcd — modelhuggingface.co—Jev-inspired 0.6B decision model on Qwen3-0.6B-Base trained with RL from right/wrong feedback so its option probabilities match accuracy; includes training loop, ablation, eval and open weights.Open models & compatible serversTraining & distillation
I built Jev for freewww.youtube.com—Builds a free Jev-style model by distilling Qwen3 0.6B from a 27B teacher on one RTX 5090, then plays Doom through ViZDoom; agreement with the teacher rose from 54% to 93%.Open models & compatible serversTraining & distillation
ifllm-learngithub.com12Research toolkit built on SemIf and MLX-LM that fine-tunes local LoRA adapters into typed decision models from labeled examples, with measured demos on support routing, out-of-scope detection and contract checklists.Open models & compatible serversTraining & distillation
Jev Distill Corpus v3huggingface.co—Typed-decision corpus of 740,957 rows in the System One noul/choice/score schema, largely synthetic scenarios across 53 domains labeled by Jev 1.13 via OpenRouter, for training small local judges.Open models & compatible serversTraining & distillation
Jev-Gate Student Bhuggingface.co—LoRA on Qwen2.5-0.5B distilled from Jev into a local memory-relevance judge that outputs a calibrated P(relevant) to decide which vector-recalled memories enter an agent context.Open models & compatible serversTraining & distillation
Jev-Gate Student B — modelhuggingface.co—LoRA on Qwen2.5-0.5B distilled from Jev into a local memory-relevance judge that outputs a calibrated P(relevant) to decide which vector-recalled memories enter an agent context.Open models & compatible serversTraining & distillation
Jev-Gate Student B — model 2huggingface.co—LoRA on Qwen2.5-0.5B distilled from Jev into a local memory-relevance judge that outputs a calibrated P(relevant) to decide which vector-recalled memories enter an agent context.Open models & compatible serversTraining & distillation
jev-qwen3.5-4b-legal-lorahuggingface.co—QLoRA adapter on Qwen3.5-4B-Base for finite-choice Chinese legal judgments that scores candidate letters in one forward pass instead of generating legal reasoning, with evaluation and inference scripts.Open models & compatible serversTraining & distillation
jev-qwen3.5-4b-legal-lora — modelhuggingface.co—QLoRA adapter on Qwen3.5-4B-Base for finite-choice Chinese legal judgments that scores candidate letters in one forward pass instead of generating legal reasoning, with evaluation and inference scripts.Open models & compatible serversTraining & distillation
jev-qwen3.5-4b-legal-lora — model 2huggingface.co—QLoRA adapter on Qwen3.5-4B-Base for finite-choice Chinese legal judgments that scores candidate letters in one forward pass instead of generating legal reasoning, with evaluation and inference scripts.Open models & compatible serversTraining & distillation
jevlite datasethuggingface.co—Synthetic set of 5,866 typed questions about 978 program states (tickets, SIEM alerts, invoices, agent transcripts, code reviews) with full teacher distributions, used to train jev-lite.Open models & compatible serversTraining & distillation
JevTunergithub.com9Training method for language models to make Jev-style closed-set decisions, scoring every candidate in one forward pass and tuning the full distribution with a tokenized Brier loss.Open models & compatible serversTraining & distillation
Jimothygithub.com24Distills Jev answers into small, fast local classifiers that run in the browser or Node.js, with calibrated probabilities and a recommended cutoff.Open models & compatible serversTraining & distillation
Jimothy — appjimothy-r63s.vercel.app—Distills Jev answers into small, fast local classifiers that run in the browser or Node.js, with calibrated probabilities and a recommended cutoff.Open models & compatible serversTraining & distillation
Jimothy — demox.com—Distills Jev answers into small, fast local classifiers that run in the browser or Node.js, with calibrated probabilities and a recommended cutoff.Open models & compatible serversTraining & distillation
Lucegithub.com7Recipe for task-specific calibrated decision models: describe the task, have an LLM teacher synthesize or annotate data, train a LoRA plus decision head on Qwen, evaluate calibration and serve typed probabilities.Open models & compatible serversTraining & distillation
minojev — docszeredy879.github.io—Library that freezes a language model and trains a small decision head to read typed, calibrated choice, boolean and score distributions from hidden states in one forward pass, reproducible on a laptop CPU.Open models & compatible serversTraining & distillation
RE Jev research projectgithub.com—Research handoff for building a Jev-like decision model, covering design criteria, architecture, a training data plan, GLiNER2 benchmark and calibration protocols; no model has been trained yet.Open models & compatible serversTraining & distillation
System One (distilled)huggingface.co—A 70.8M DeBERTa-v3-xsmall cross-encoder for the Jev /v1/systemone schema, trained on Claude Haiku 4.5's answers to compare distillation against gold labels.Open models & compatible serversTraining & distillation
System One training pairshuggingface.co—Training bundle for a DeBERTa System One cross-encoder: gold premise-hypothesis pairs, Claude Haiku 4.5 distillation labels, fixed Banking77 and tool-selection splits, and the fine-tuning code.Open models & compatible serversTraining & distillation
System One training pairs — projectgithub.com—Training bundle for a DeBERTa System One cross-encoder: gold premise-hypothesis pairs, Claude Haiku 4.5 distillation labels, fixed Banking77 and tool-selection splits, and the fine-tuning code.Open models & compatible serversTraining & distillation
system-one-270m-datahuggingface.co—Synthetic set of 25,002 typed decisions with soft target distributions over caller-supplied options, rendered as lettered prompts, for training the open system-one-270m model.Open models & compatible serversTraining & distillation
tev1github.com191Data recipe and training code that fine-tune Qwen3.5-4B into an open-weight decision model on Together AI.Open models & compatible serversTraining & distillation
Train your own Jev-like modelx.com—Trains a Jev-style judgment model with RLCD on Qwen3.5-4B plus a rank-8 LoRA (32K questions, 9 GPU hours, about $42), tying Jev on the 120-question JevBench.Open models & compatible serversTraining & distillation
Train your own Jev-like model — modelhuggingface.co—Trains a Jev-style judgment model with RLCD on Qwen3.5-4B plus a rank-8 LoRA (32K questions, 9 GPU hours, about $42), tying Jev on the 120-question JevBench.Open models & compatible serversTraining & distillation
Train your own Jev-like model — projecthuggingface.co—Trains a Jev-style judgment model with RLCD on Qwen3.5-4B plus a rank-8 LoRA (32K questions, 9 GPU hours, about $42), tying Jev on the 120-question JevBench.Open models & compatible serversTraining & distillation
Compatible servers30 of 30 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
AINode System One APIgithub.com—System One-compatible decide API in AINode, a self-hosted AI platform for NVIDIA GPUs, benchmarked on 110 labeled decisions: hosted Jev scored 0.964 accuracy, Brier 0.024 and 0 of 80 wrong at a 0.9 gate vs a local 35B chat model.Open models & compatible serversCompatible servers
AINode System One API — appainode.dev—System One-compatible decide API in AINode, a self-hosted AI platform for NVIDIA GPUs, benchmarked on 110 labeled decisions: hosted Jev scored 0.964 accuracy, Brier 0.024 and 0 of 80 wrong at a 0.9 gate vs a local 35B chat model.Open models & compatible serversCompatible servers
AINode System One API — benchgithub.com—System One-compatible decide API in AINode, a self-hosted AI platform for NVIDIA GPUs, benchmarked on 110 labeled decisions: hosted Jev scored 0.964 accuracy, Brier 0.024 and 0 of 80 wrong at a 0.9 gate vs a local 35B chat model.Open models & compatible serversCompatible servers
AINode System One API — repogithub.com—System One-compatible decide API in AINode, a self-hosted AI platform for NVIDIA GPUs, benchmarked on 110 labeled decisions: hosted Jev scored 0.964 accuracy, Brier 0.024 and 0 of 80 wrong at a 0.9 gate vs a local 35B chat model.Open models & compatible serversCompatible servers
arbitergithub.com19Serving layer for typed-decision models such as Laya or your own on NVIDIA GPUs or Apple Silicon behind a Jev-compatible API, with coding-agent integrations and examples like a PR risk gate.Open models & compatible serversCompatible servers
djev-rungithub.com32Deployment recipe that serves DiffusionGemma-Jev on Google Cloud Run with an RTX PRO 6000 Blackwell GPU behind a TypeSafe-compatible API, plus a zero-dependency Snake demo calling it from the browser.Open models & compatible serversCompatible servers
djev-sparkgithub.com169Container recipe that serves DiffusionGemma 26B-A4B in NVFP4 on a DGX Spark behind Jev's /v1/systemone API, pairing patched vLLM structured reads with a separate decision server.Open models & compatible serversCompatible servers
endpointgithub.com560razorback16/openjev — Open, Jev-compatible System One decision server on DiffusionGemma · endpoint · PythonOpen models & compatible serversCompatible servers
endpointgithub.com1zerodegress/jevinf — Jev-like model inference engine + Jev-compatible API · endpoint · PythonOpen models & compatible serversCompatible servers
Jeff — demox.com—Compatible server built on a 400M-parameter encoder, with benchmark comparisons against Jev.Open models & compatible serversCompatible servers
Jev Localgithub.com20Unofficial local /v1/systemone server for text, JSON and images that answers Choice, Score and Noul via first-token logprobs of LFM2.5 or Sarashina models, or a fine-tuned Japanese ModernBERT cross-encoder.Open models & compatible serversCompatible servers
Jev Local — modelhuggingface.co—Unofficial local /v1/systemone server for text, JSON and images that answers Choice, Score and Noul via first-token logprobs of LFM2.5 or Sarashina models, or a fine-tuned Japanese ModernBERT cross-encoder.Open models & compatible serversCompatible servers
Jev Local — model 2huggingface.co—Unofficial local /v1/systemone server for text, JSON and images that answers Choice, Score and Noul via first-token logprobs of LFM2.5 or Sarashina models, or a fine-tuned Japanese ModernBERT cross-encoder.Open models & compatible serversCompatible servers
jevinfgithub.com1Inference engine for Jev-like decision models that runs each candidate path as segmented forwards with prefix reuse and serves the Jev wire contract, measured 2.57x faster on MPS with NanoJev.Open models & compatible serversCompatible servers
local-jevgithub.com13Offline server that speaks TypeSafe's System One wire format so the official SDK works unchanged, answering typed questions with small open models; reports 80.5% on JevBench's 231 public items.Open models & compatible serversCompatible servers
LocalJevgithub.com707Bun server from GitHub Next that exposes a local Jev-compatible /v1/systemone API backed by DiffusionGemma, prompting for JSON probabilities instead of reading logits, with a 1,200-request bake-off.Open models & compatible serversCompatible servers
LocalJev — demox.com—Bun server from GitHub Next that exposes a local Jev-compatible /v1/systemone API backed by DiffusionGemma, prompting for JSON probabilities instead of reading logits, with a 1,200-request bake-off.Open models & compatible serversCompatible servers
LocalJev — discussionnews.ycombinator.com—Bun server from GitHub Next that exposes a local Jev-compatible /v1/systemone API backed by DiffusionGemma, prompting for JSON probabilities instead of reading logits, with a 1,200-request bake-off.Open models & compatible serversCompatible servers
ollayagithub.com1,059Pulls and serves open decision models such as Laya, kev, and JevK5 behind a TypeSafe-compatible local endpoint, the way Ollama serves LLMs.Open models & compatible serversCompatible servers
openjev-sglang — discussionwww.reddit.com—Compatible endpoint that serves an open mixture-of-experts model on SGLang, with BoolQ and MMLU-Pro comparisons.Open models & compatible serversCompatible servers
PCD Servergithub.com5Native C++ REST server for Parallel Constrained Decoding with local GGUF models: give it text and bounded fields and it picks one allowed value per field, returning JSON plus each decision's probabilities, Jev-style.Open models & compatible serversCompatible servers
Rizzo Flowgithub.com780Local open-model decision server that reads answer-token probabilities instead of generating text, behind a Jev-compatible Choice, Score and Noul HTTP API plus a numeric primitive; no quality parity with Jev is claimed.Open models & compatible serversCompatible servers
Rizzo Flow — siterizzo-ai-academy.github.io—Local open-model decision server that reads answer-token probabilities instead of generating text, behind a Jev-compatible Choice, Score and Noul HTTP API plus a numeric primitive; no quality parity with Jev is claimed.Open models & compatible serversCompatible servers
snapjudgegithub.com14TypeSafe-compatible server that reads Choice, Score, and Noul probabilities from the logits of local Qwen models on Apple Silicon via MLX; Qwen3.8-27B reaches 81.2% agreement with Jev on 20 public cases.Open models & compatible serversCompatible servers
snapjudge — discussionwww.reddit.com—TypeSafe-compatible server that reads Choice, Score, and Noul probabilities from the logits of local Qwen models on Apple Silicon via MLX; Qwen3.8-27B reaches 81.2% agreement with Jev on 20 public cases.Open models & compatible serversCompatible servers
SparkStation decision modelsgithub.com—Model fleet manager for NVIDIA DGX Spark that serves decision models such as Reflex over a TypeSafe-Jev-compatible POST /v1/systemone endpoint next to its OpenAI-compatible gateway.Open models & compatible serversCompatible servers
SparkStation decision models — repogithub.com—Model fleet manager for NVIDIA DGX Spark that serves decision models such as Reflex over a TypeSafe-Jev-compatible POST /v1/systemone endpoint next to its OpenAI-compatible gateway.Open models & compatible serversCompatible servers
Splashgithub.com579Local Apple silicon inference engine for coding agents that also serves a TypeSafe-compatible /v1/systemone endpoint, answering noul, choice and score questions from logits and working with the official SDK.Open models & compatible serversCompatible servers
Splash — articleinco.ai—Local Apple silicon inference engine for coding agents that also serves a TypeSafe-compatible /v1/systemone endpoint, answering noul, choice and score questions from logits and working with the official SDK.Open models & compatible serversCompatible servers
typesafe-localgithub.com8Local MLX server with TypeSafe's API shape that encodes a document once, appends each question as a short suffix and reads logits instead of generating; four questions take ~200 ms on an M4 Pro.Open models & compatible serversCompatible servers
Compatibility adapters6 of 6 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
AnyJevgithub.com997Library that turns any transformers or vLLM model into a Jev-style Choice, Score, and Noul decision model from one prefill with no training, fixing option-order flips via cyclic-shift marginalization.Open models & compatible serversCompatibility adapters
AnyJev — pypipypi.org—Library that turns any transformers or vLLM model into a Jev-style Choice, Score, and Noul decision model from one prefill with no training, fixing option-order flips via cyclic-shift marginalization.Open models & compatible serversCompatibility adapters
deepseek-v4.1-flash-jevx.com—Endpoint that makes DeepSeek V4.1 Flash behave like Jev by using SGLang's scoring API to return probabilities over a fixed answer set instead of generated text.Open models & compatible serversCompatibility adapters
Jev-Compatiblegithub.com6HTTP gateway that turns an existing SGLang or vLLM deployment into a Jev-compatible decision service by scoring candidate tokens for noul, choice, and score questions, with no training or model changes.Open models & compatible serversCompatibility adapters
verdictgithub.com5Self-hosted server that turns any llama-server or llama-swap endpoint into a Jev-compatible System One API by reading the probability on each option label in one forward pass, so Jev clients only change the base URL.Open models & compatible serversCompatibility adapters
verdict — demonews.ycombinator.com—Self-hosted server that turns any llama-server or llama-swap endpoint into a Jev-compatible System One API by reading the probability on each option label in one forward pass, so Jev clients only change the base URL.Open models & compatible serversCompatibility adapters
More projects & source code217 of 217 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
@receptron/layagithub.com205Node.js and TypeScript package that runs Laya, the open Jev-compatible decision model by Convai Innovations, through ONNX Runtime and returns choice, score, and noul answers in one forward pass.Open models & compatible serversMore projects & source code
@ruvector/typesafegithub.com—Local, network-free typed decisions over sentence embeddings that accept Jev's /v1/systemone request and response shape, returning calibrated confidence and an abstain mass.Open models & compatible serversMore projects & source code
AgentKthx JEV API modegithub.com—Mode in the AgentKthx local-first agent framework that emulates Jev's System One decision shape with any chat LLM (Ollama, ZAI, OpenRouter, llama-server) by wrapping a constrained prompt and parsing a decision envelope.Open models & compatible serversMore projects & source code
AgentKthx JEV API mode — repogithub.com—Mode in the AgentKthx local-first agent framework that emulates Jev's System One decision shape with any chat LLM (Ollama, ZAI, OpenRouter, llama-server) by wrapping a constrained prompt and parsing a decision envelope.Open models & compatible serversMore projects & source code
any2jevgithub.com5Converter that trains any Hugging Face causal LM into a Jev-style decision model answering typed questions in one forward pass with calibrated probabilities and zero generated tokens, demoed on Qwen3-0.6B.Open models & compatible serversMore projects & source code
AnyJev — READMEgithub.com997AnyJev presents a research/runtime path for turning existing language models into Jev-style typed decision models; the README claims are not independently reproduced here.Open models & compatible serversMore projects & source code
Bespoke Nimblegithub.com1,523Open data, recipe, and Qwen3.5-9B LoRA model for Jev-style typed decisions, trained on contrastive negatives made by flipping one fact; reports 90.1% agreement with Jev versus 66.4% for base Qwen.Open models & compatible serversMore projects & source code
bivio — READMEgithub.com0README presents Bivio as a model inspired by Jev and says it speaks the same API; that compatibility claim is not independently verified.Open models & compatible serversMore projects & source code
bokuweb/grandegithub.com1A System One style decision model runtime in Rust.Open models & compatible serversMore projects & source code
bonzi-8b-ternary-v1-jev — projectgithub.com—Recipe and measurements (no weights) for running Ternary Bonsai 1 8B GGUF as a Jev-style typed decision function that returns a probability per option from one forward pass.Open models & compatible serversMore projects & source code
Brain DOOM decision policy — game-servergithub.com—Local Jev-style policy that plays DOOM in real time: a frozen MiniLM encoder reads the observation text and a small trained head picks actions, at 58.3 ms per decision.Open models & compatible serversMore projects & source code
Brain DOOM decision policy — repogithub.com—Local Jev-style policy that plays DOOM in real time: a frozen MiniLM encoder reads the observation text and a small trained head picks actions, at 58.3 ms per decision.Open models & compatible serversMore projects & source code
browser-jevgithub.com1Run Jev-like decisions in your browser with the Prompt API.Open models & compatible serversMore projects & source code
browser-jev — READMEgithub.com1A browser-local Jev-format alternative uses Chrome's model for yes/no, choices and scores; it does not call TypeSafe Jev, and confidence is not calibrated correctness.Open models & compatible serversMore projects & source code
burgerwdev/what-is-jevgithub.com1Jev notes: fourteen decisions I ran through itOpen models & compatible serversMore projects & source code
ChatJevgithub.com7Small experiment that puts Jev in an autoregressive loop, offering candidate next tokens as Choice options to make the non-generative model produce text one token at a time.Open models & compatible serversMore projects & source code
choosekitgithub.com18Library that scores a finite set of choices with llama.cpp, Ollama, or OpenRouter models and returns typed decisions with probability distributions, plus an MCP server and a SuperGPQA benchmark against Jev 1.13.Open models & compatible serversMore projects & source code
CLM — READMEgithub.com2,658The CLM project presents a TypeSafe-compatible API and claims performance parity with Jev on several tasks; these README results were not independently verified.Open models & compatible serversMore projects & source code
corasan/arjevgithub.com1Argent + JevOpen models & compatible serversMore projects & source code
cu-Jevgithub.com4C/CUDA decision engine that turns open Qwen3.5 checkpoints (0.8B to 9B) into a Jev-compatible /v1/systemone API with hand-written kernels and no PyTorch at runtime, plus reproducible benchmarks.Open models & compatible serversMore projects & source code
cu-Jev — READMEgithub.com4cu-Jev is a separate CUDA/Qwen3.5 typed-decision engine; its README claims a Jev-compatible API, which is not independently verified here.Open models & compatible serversMore projects & source code
cua-s1-forms — repogithub.com≈26,700Tiny Jev-like option scorer (706,048 parameters, 2.8 MB) that rates fill, check, click or skip for every form field in one parallel pass, built as the decision layer for cua-driver.Open models & compatible serversMore projects & source code
decidagithub.com0system one server and test bench for typed decisionsOpen models & compatible serversMore projects & source code
decida — READMEgithub.com0Local System One runtime serves typed decision models through REST and MCP, with hosted Jev available as an optional comparison backend.Open models & compatible serversMore projects & source code
Decidergithub.com1,023One-pass typed decisions with calibrated probabilities (System One style model), fine-tuned from Qwen3.5-2B.Open models & compatible serversMore projects & source code
decider — READMEgithub.com1,023Decider is an independent typed-decision model positioned as a System One/Jev-class alternative; its README explicitly says it was not distilled from Jev.Open models & compatible serversMore projects & source code
decima — READMEgithub.com0Independent compact Jev-style decision model implements choose/score/verify/rank without text generation, with ONNX CPU/browser deployment and multilingual option inputs.Open models & compatible serversMore projects & source code
Decisgithub.com7Self-hosted, Jev-compatible decision-model API — one /v1/systemone endpoint, open weights (Laya, kev), one Docker image per engine.Open models & compatible serversMore projects & source code
Decis — READMEgithub.com7A self-hostable inference server for open decision models implementing a TypeSafe-style System One interface for existing SDK clients. It is an alternative serving layer, not a hosted Jev model.Open models & compatible serversMore projects & source code
deep-diver/mini-jevgithub.com3mini-jev: Google Gemma 3 270M 기반 Jev 호환 초고속 의사결정 엔진Open models & compatible serversMore projects & source code
DeepSeek-V4.1 A100 Jev Modegithub.com54Custom DeepSeek-V4.1-Flash runtime for A100 GPUs with a Jev Mode that scores candidates in parallel instead of decoding JSON, reporting up to 2,854x speedup on 30-field extraction in 24 ms.Open models & compatible serversMore projects & source code
dev-0.4bgithub.com19Open 399M-parameter decision model on ModernBERT-large for developer tools and coding agents, answering routing, yes/no and 1-5 rating questions in one forward pass (~28 ms on Apple Silicon).Open models & compatible serversMore projects & source code
djevgithub.com110Example server from a vLLM pull request that gets Jev-style structured decisions from DiffusionGemma by seeding the diffusion canvas with the answer template and reading slot distributions after one denoise step.Open models & compatible serversMore projects & source code
endpointgithub.com20Argos1111/jev_local — Replicating Jev with a local LLM · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com14GPT-AGI/OpenJev — Jev-compatible System 开源Jev · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com2Saik0s/diffusiongemma-jev-macos — Local JEV-style decisions with DiffusionGemma on Apple Silicon, with benchmarks and coding-agent examples. · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com2andrueandersoncs/visual-jev — Image-native typed decisions with shared visual encoding and Qwen3-VL · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com121daseinlabs/open-jev — Open Jev implementation with custom finetuning · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com57deepanwadhwa/OpenDecision — OpenDecision is an open-source semantic decision engine like typesafe's jev. · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com3fritzprix/systemone-lite — Toy local System One–style decision API (Jev-shaped). Not affiliated with TypeSafe. · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com13genai-craft/openvons — openvons (open-Jev): 有限選択肢に確率で答える判断層 — テキスト / 画像 / 日本語音声コマンド · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com8,135jaredpalmer/kev — tiny Jev-like model built on top of Qwen2.5-0.5B you can train and run on your MacBook · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com3kvnloo/z0intelligence — Can we run something like Jev on a 3090 at home? · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com276logan-markewich/jeff — A self-hosted drop-in replacement for TypeSafe's jev, powered by GliFormer. · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com3rorshopping/jev-browser-local — Run jev-browser on a fully local JEV-style decision engine (no cloud API). Warm-browser fork, VRAM guard, measured benchmarks, run traces. · endpoint · PythonOpen models & compatible serversMore projects & source code
endpointgithub.com105zwliJay/jev-forge — An open training and inference stack for Jev-style decision models. Train models to score dynamic candidate branches from a shared prefix, with support for high-cardinality choice, calibration, and fast batched inference. · endpoint · PythonOpen models & compatible serversMore projects & source code
Eugeniuszgithub.com9MIT-licensed C++17 library with a C ABI that evaluates text locally and returns a choice, rubric score or truth probability, as an independent alternative to the Jev decision interface for C, C#, Python, Unity and Unreal.Open models & compatible serversMore projects & source code
ggmlc Laya enginegithub.com—Example in the ggmlc neural-network compiler that compiles the open Jev alternative Laya into a standalone C++ decision engine with GGUF weights and a TypeSafe-compatible /v1/systemone server; ~25 ms per decision on an RTX 4050 Laptop.Open models & compatible serversMore projects & source code
ggmlc Laya engine — repogithub.com—Example in the ggmlc neural-network compiler that compiles the open Jev alternative Laya into a standalone C++ decision engine with GGUF weights and a TypeSafe-compatible /v1/systemone server; ~25 ms per decision on an RTX 4050 Laptop.Open models & compatible serversMore projects & source code
Glancegithub.com7Harness for asking an open vision-language model (Qwen3-VL-4B by default) typed yes/no, pick-one and rating questions about an image, reading calibrated probabilities from one forward pass, with Jev-style request shapes.Open models & compatible serversMore projects & source code
hearimgithub.com4Go gateway that serves the Jev /v1/systemone contract on ordinary LLM backends like Ollama, llama.cpp and vLLM by scoring each question as a single-token choice from logprobs.Open models & compatible serversMore projects & source code
hedos System One gatewaygithub.com—Headless local model engine whose gateway exposes TypeSafe's POST /v1/systemone, so a TypeSafe SDK pointed at it gets answers from a local judge-capable model.Open models & compatible serversMore projects & source code
hedos System One gateway — repogithub.com—Headless local model engine whose gateway exposes TypeSafe's POST /v1/systemone, so a TypeSafe SDK pointed at it gets answers from a local judge-capable model.Open models & compatible serversMore projects & source code
HyprStream decision servicegithub.com—Self-improving AI runtime building its own System One decision service, with a decision head and a /v1/systemone facade that the stock TypeSafe Python and JS SDKs run against.Open models & compatible serversMore projects & source code
imajev — READMEgithub.com186An independent local decision-model implementation that extends Jev-style typed answers to images. The README describes running on the user’s own hardware and an image-extended Jev request contract; it is listed as a related alternative, not as TypeSafe Jev or independently verified API compatibility or benchmark superiority.Open models & compatible serversMore projects & source code
INSTRUCT_JEVgithub.com1Instruction corpus of 119 Choice, Noul and Score rows compiled from TypeSafe's public Jev documentation, published as a HuggingFace dataset.Open models & compatible serversMore projects & source code
Intel-Phi-Jevgithub.com1XKEYSCORE for Jev: TypeSafe System One judgments (noul/choice/score) served locally, with the Xeon Phi cards doing the model's matrix workOpen models & compatible serversMore projects & source code
Intel-Phi-Jev — READMEgithub.com1Local System One implementation reads an LLM’s next-token probabilities to answer Jev-style typed questions, rather than running hosted Jev.Open models & compatible serversMore projects & source code
jeff by logan-markewichgithub.com276Self-hosted System One API on the 400M GLiFormer, with benchmarks that say where it trails Jev.Open models & compatible serversMore projects & source code
jeff-inkgithub.com0Type anything, the story stays authored. Jev maps free text to Ink choices.Open models & compatible serversMore projects & source code
jeff-ink — READMEgithub.com0An Ink-story wrapper maps each typed player line to one of the author’s existing choices through Jev; the README says it generates no story text.Open models & compatible serversMore projects & source code
Jev mode for llama.cpp — playgroundgithub.com7Fork of llama.cpp adding a llama-server endpoint that takes instructions, a schema, and state and returns every field with a confidence score; 17.3 ms per decision in bulk on a 2 GB model.Open models & compatible serversMore projects & source code
Jev mode for llama.cpp — projectgithub.com—Fork of llama.cpp adding a llama-server endpoint that takes instructions, a schema, and state and returns every field with a confidence score; 17.3 ms per decision in bulk on a 2 GB model.Open models & compatible serversMore projects & source code
Jev mode for llama.cpp — repogithub.com—Fork of llama.cpp adding a llama-server endpoint that takes instructions, a schema, and state and returns every field with a confidence score; 17.3 ms per decision in bulk on a 2 GB model.Open models & compatible serversMore projects & source code
Jev Quantumgithub.com22Jev-protocol-compatible random baseline in Rust that speaks noul, choice and score but ignores the prompt and answers from a fast pseudo-random generator, for use as a mock or a lower bound in agent-routing evals.Open models & compatible serversMore projects & source code
Jev Visualgithub.com303Educational Jev-like visual inference on Apple Silicon: Qwen3.5-0.8B on MLX answers choice, yes/no and score questions about one image from shared context and logits, with sorting-factory, Breakout and gesture demos.Open models & compatible serversMore projects & source code
Jev vs OpenJev FPS — openjevgithub.com335First-person shooter in the browser where Jev fights OpenJev, an open Jev-compatible API running Qwen on SGLang.Open models & compatible serversMore projects & source code
jev-adrankgithub.com0Real-time ad ranking and creative review on a typed evaluation model — one call prices a whole auction, no trained CTR model, no logged clicks.Open models & compatible serversMore projects & source code
jev-adrank — READMEgithub.com0The ad-ranking experiment exposes jevBid and a separate JEV_REC_OFFLINE=1 keyword-stub mode.Open models & compatible serversMore projects & source code
jev-agent-labgithub.com—Reproducible Runpod deployment of the open Jev-style SemIf model alongside an official Jev client, with agent decision cases, latency measurements and a Jev vs SemIf comparison report.Open models & compatible serversMore projects & source code
JEV-CPUgithub.com12CPU port of SemIf with a web UI that reads typed option probabilities from Qwen3-0.6B logits in about 1 s, covering eight domains such as moderation, incident severity and credit risk.Open models & compatible serversMore projects & source code
jev-decision-labgithub.com0A self-hosted, bilingual playground for Jev’s Choice, Noul, and Score judgments.Open models & compatible serversMore projects & source code
jev-decision-lab — READMEgithub.com0Unofficial Jev playground defines Choice/Noul/Score in forms and compares single, combined or concurrent calls with detailed response metrics; a deployer-provided server credential replaces visitor key entry.Open models & compatible serversMore projects & source code
jev-lite — projectgithub.com—QLoRA adapter that turns Gemma 4 E4B into a System One decision model, reading answers from the logits at the option letters so it cannot answer outside the given options.Open models & compatible serversMore projects & source code
jev-llmgithub.com1Fake autoregressive language model powered by TypeSafe JevOpen models & compatible serversMore projects & source code
jev-my-bro governance dataset — projectgithub.com—English/Thai dataset of 8,508 cases and 34,032 typed decisions on whether an agent operation should execute, ask the user or be rejected, needs review, is prohibited, and how risky it is.Open models & compatible serversMore projects & source code
jev-pakkiogithub.com0TypeSafe Jev ecosystem repository.Open models & compatible serversMore projects & source code
jev-pakkio — READMEgithub.com0Local Jev-style option scorer uses Gemma 3 4B on MLX or PyTorch, ranking predefined options in a batched forward pass rather than generating text.Open models & compatible serversMore projects & source code
jev-patternsgithub.com0Typed distribution patterns for Jev responses: descriptive shapes, composable predicates, and exhaustive matching.Open models & compatible serversMore projects & source code
jev-patterns — READMEgithub.com0Independent TypeScript library analyzes Jev response distributions into typed concentration, alternatives and rankings, leaving policy decisions to application code and making no network calls.Open models & compatible serversMore projects & source code
jev-vllmgithub.com8TypeSafe Jev ecosystem repository.Open models & compatible serversMore projects & source code
jev-vllm — READMEgithub.com8A self-hosted open-weight decision service returning typed probabilities through a Jev-style System One API. Listed as a related alternative, not TypeSafe’s hosted Jev or verified API compatibility.Open models & compatible serversMore projects & source code
jevextractgithub.com1Grounded information extraction that cannot hallucinate: code proposes spans, Jev decides. An open-source alternative to LangExtract, with a bilingual benchmark and paper.Open models & compatible serversMore projects & source code
jevextract — READMEgithub.com1A grounded extraction tool where code proposes source spans and TypeSafe Jev classifies candidates or none. Verbatim spans constrain output, but classification can still be wrong.Open models & compatible serversMore projects & source code
jevgraph — READMEgithub.com32A schema-guided document-to-graph pipeline uses Jev to choose candidate relations after local candidate blocking.Open models & compatible serversMore projects & source code
jevifygithub.com3Experimental inference engine that makes an existing LLM (Qwen2.5-1.5B-Instruct by default) answer Jev-style classification, yes/no and rubric questions by building JSON from model scores; probabilities are uncalibrated.Open models & compatible serversMore projects & source code
jevify-gemma4-26b-a4b — projectgithub.com—Gemma 4 26B-A4B fine-tuned (LoRA, merged) to give honest probabilities on typed questions; the model behind jevify, a local Jev-compatible /v1/systemone decision API.Open models & compatible serversMore projects & source code
jevify-qwen3.5-4b — projectgithub.com4Jevify decision heads for Qwen3.5-4B that read a state, answer choice, score and noul questions and return calibrated distributions, shipping only the heads added on top of the base model.Open models & compatible serversMore projects & source code
jevk5 — READMEgithub.com128JevK5 describes an independent typed-decision model that returns a probability for each option in one forward pass.Open models & compatible serversMore projects & source code
jevmlxgithub.com69Local Jev-style decision layer for Apple Silicon that scores every allowed option of boolean, enum and multi-select fields from MLX model logits in one prefill and returns schema-valid JSON with per-field probabilities.Open models & compatible serversMore projects & source code
jevons-lfm25-1.2b-systemone — projectgithub.com—Seed LoRA on LFM2.5-1.2B for jevons, a local System One server that scores allowed outcomes from logits and assembles choice, probabilities, confidence, noul and score in code.Open models & compatible serversMore projects & source code
jevos — LICENSEgithub.com1,154Independent, MIT-licensed local decision implementation that returns probabilities for yes/no questions. Its published jevos model currently supports Noul only; this listing does not certify full Jev compatibility or the author’s benchmark results.Open models & compatible serversMore projects & source code
jevos — READMEgithub.com1,154Independent, MIT-licensed local decision implementation that returns probabilities for yes/no questions. Its published jevos model currently supports Noul only; this listing does not certify full Jev compatibility or the author’s benchmark results.Open models & compatible serversMore projects & source code
jevtppgithub.com5Jev-like decisions. Strong C++ types.Open models & compatible serversMore projects & source code
jevtpp — READMEgithub.com5JevT++ is an independent C++20 library for model-backed routing, classification, and scoring, not the proprietary Jev model or an official TypeSafe SDK.Open models & compatible serversMore projects & source code
KaLM-Jevgithub.com28Local Choice, Score, and Noul judgment service built on KaLM-Reranker-V1-R2 checkpoints in Nano, Small, and Large sizes, returning structured judgments without generating answer text.Open models & compatible serversMore projects & source code
Kapteenigithub.com0A Jev-compatible System One decision modelOpen models & compatible serversMore projects & source code
Kapteeni — READMEgithub.com0Local System One decision-model implementation adapts Qwen3-4B with LoRA and a readout layer to return typed answers with probabilities and confidence.Open models & compatible serversMore projects & source code
kevgithub.com8,135LoRA adapter and readout head on Qwen2.5-0.5B that answers many typed questions in one prefill; trains in under two hours on a MacBook, held-out ECE 0.065, speaks the TypeSafe wire format.Open models & compatible serversMore projects & source code
kev — READMEgithub.com8,135Kev is a Qwen3.5-based family of Jev-like decision models that can be trained locally; its README says Jev outputs were not used for training.Open models & compatible serversMore projects & source code
kevMacgithub.com18Mac app that runs the open Kev decision models locally on Apple Silicon: paste a document, build yes/no, choice or score questions as forms and see calibrated probability bars; kev-0.6b answers in ~0.2 s on an M4.Open models & compatible serversMore projects & source code
kvnloo/z0intelligencegithub.com3Can we run something like Jev on a 3090 at home?Open models & compatible serversMore projects & source code
Laya Gogithub.com3Go server with an admin dashboard, agent CLI and MCP tool for Laya, open System 1 decision models that score structured features and return ranked probabilities without generating text.Open models & compatible serversMore projects & source code
Laya MCPgithub.com5MCP server and agent plugin for Claude Code, Copilot CLI and Cursor that runs the open Laya decision model locally and gives agents five tools returning a label, probability or score in one forward pass.Open models & compatible serversMore projects & source code
Laya MPSgithub.com5Runs the open Laya model for Jev-style typed decisions on Apple Silicon via PyTorch MPS, at about 32 ms median latency and 2.1 GiB of RAM on an M5 Pro.Open models & compatible serversMore projects & source code
Laya playgroundgithub.com71Local website with games, a benchmark and an agent skill for the open-source Laya decision model, comparing it with hosted Jev on the same 500 labelled examples.Open models & compatible serversMore projects & source code
laya-guardrailsgithub.com0Guardrails de input, tool call y output para agentes de IA con un modelo System One tipo Jev (laya-pt-es-typed) autoalojado en una NVIDIA L4: 5 ms por check frente a 140 ms de un LLM-as-a-judge.Open models & compatible serversMore projects & source code
laya-guardrails — READMEgithub.com0Self-hosted agent guardrails use a Laya decision model for input, tool-call and output checks, following a Jev-like typed System One approach rather than hosted Jev.Open models & compatible serversMore projects & source code
Laya-MLXgithub.com6,673Native MLX runtime for the open Laya typed-decision checkpoints on Apple Silicon: 13.4 ms median per short English decision and 7.4 ms with the multilingual checkpoint, with no PyTorch or cloud API; demoed on Snake.Open models & compatible serversMore projects & source code
laya-mlx — READMEgithub.com6,673The README presents an independent MLX runtime for open-weight typed decisions on Apple Silicon, relevant as a local Jev-adjacent alternative; it does not claim to be TypeSafe Jev.Open models & compatible serversMore projects & source code
laya-servergithub.com1A proof of concept implementation of TypeSafe's Jev API, backed by local Laya checkpoints.Open models & compatible serversMore projects & source code
laya-server — READMEgithub.com1The README presents a proof of concept for serving local Laya checkpoints behind a Jev API-style endpoint; it does not establish equivalence to Jev.Open models & compatible serversMore projects & source code
laya.cppgithub.com40Standalone C++ inference for the Jev-style Laya typed-decision models using ggml, with optimized CUDA kernels for NVIDIA RTX GPUs and a Vulkan FP32 backend, covering the english, multilingual and typed-decisions checkpoints.Open models & compatible serversMore projects & source code
laya_exgithub.com13Elixir library that provides a native Nx and Bumblebee runtime for the open Laya decision model by Convai Innovations, downloading the official checkpoint on first load.Open models & compatible serversMore projects & source code
layaForWeb — projectgithub.com—Runs the open 421M-parameter Laya decision model inside a browser tab via ONNX Runtime Web and WebAssembly, shown next to Jev's playground; the browser build matched PyTorch's top answer on all 14 questions.Open models & compatible serversMore projects & source code
layaForWeb — repogithub.com—Runs the open 421M-parameter Laya decision model inside a browser tab via ONNX Runtime Web and WebAssembly, shown next to Jev's playground; the browser build matched PyTorch's top answer on all 14 questions.Open models & compatible serversMore projects & source code
LCT — READMEgithub.com5An independent Jev-style System One model research project using recurrent computation trajectories for decision confidence and documenting a typed HTTP interface. Performance and calibration claims are unverified here.Open models & compatible serversMore projects & source code
levgithub.com9Jev-style decision model based on LFM2.5-350MOpen models & compatible serversMore projects & source code
lev — READMEgithub.com43An independent System One decision model built with a Qwen3.5-4B LoRA adapter. It answers typed yes/no, choice and score questions from logits and documents a TypeSafe-style /v1/systemone interface. Listed as a related alternative, not as TypeSafe Jev or independently tested API compatibility or benchmark performance.Open models & compatible serversMore projects & source code
lev — READMEgithub.com9Lev is an independent Jev-style prototype based on LiquidAI LFM2.5-350M, returning probabilities for typed questions without text generation.Open models & compatible serversMore projects & source code
LitJevgithub.com45Hypothesis-based reproduction of the Jev decision layer that turns off-the-shelf Qwen checkpoints into typed Choice, Score and Noul endpoints by reading option scores from the output head, with an API and browser front end.Open models & compatible serversMore projects & source code
llama-cpp-system-one — READMEgithub.com0This independent Rust project implements the System One API using DiffusionGemma and llama.cpp; it is not TypeSafe Jev itself.Open models & compatible serversMore projects & source code
LLM2Jevgithub.com381Toolkit that turns local language models into Jev-style decision engines by reading Choice, Score and Noul answers from prefill logits via Transformers or SGLang, with a /v1/systemone endpoint and image input.Open models & compatible serversMore projects & source code
lmr-rsgithub.com0TypeSafe Jev ecosystem repository.Open models & compatible serversMore projects & source code
lmr-rs — READMEgithub.com0lmr-rs is a Rust inference server for local Laya and other open-weight models. Its README describes serving requests in the same /v1/systemone shape as TypeSafe; this is a documented interface claim, not independently tested compatibility with TypeSafe Jev.Open models & compatible serversMore projects & source code
Local Decision Playgroundgithub.com5Dependency-free web app, in Japanese, for trying a local vLLM model as a Jev-inspired classifier: it restricts generation to the declared options and normalizes their logprobs into relative probabilities, including for images.Open models & compatible serversMore projects & source code
local-judgegithub.com0TypeSafe Jev ecosystem repository.Open models & compatible serversMore projects & source code
local-judge — READMEgithub.com0Decision engine implements a documented Jev request/response subset using Ollama by default or a configured chat-model endpoint, without certifying full Jev equivalence.Open models & compatible serversMore projects & source code
local-system-one-student — projectgithub.com—A 149M ModernBERT encoder distilled from local LLMs that reproduces Jev's interface for one fixed schema, a coding agent's next step, returning tool choice, urgency and destructive-action flags in 19 ms.Open models & compatible serversMore projects & source code
maclayagithub.com0Run Laya typed-decision models locally with one command on Apple Silicon behind a Jev-compatible API, with a dashboard and playgroundOpen models & compatible serversMore projects & source code
maclaya — READMEgithub.com0Runs Laya typed-decision models on Apple Silicon through laya-mlx and exposes documented Jev-shaped endpoints, with a request dashboard, logs and playground; it is a local alternative.Open models & compatible serversMore projects & source code
mini-Jevgithub.com57Pre-registered experiment showing a Jev-style typed-decision interface on frozen Qwen3-4B by turning schema fields into lettered questions and reading option-letter logits in one pass, compared with grammar-constrained JSON.Open models & compatible serversMore projects & source code
mlx-omarchy Laya servinggithub.com—MLX GPU backend for Apple Silicon Linux that ports and serves the open Laya typed-decision model through its own decisions endpoint, qualified on an M1 Max with 6/6 numerical gates.Open models & compatible serversMore projects & source code
mlx-omarchy Laya serving — demogithub.com—MLX GPU backend for Apple Silicon Linux that ports and serves the open Laya typed-decision model through its own decisions endpoint, qualified on an M1 Max with 6/6 numerical gates.Open models & compatible serversMore projects & source code
mlx-omarchy Laya serving — docsgithub.com—MLX GPU backend for Apple Silicon Linux that ports and serves the open Laya typed-decision model through its own decisions endpoint, qualified on an M1 Max with 6/6 numerical gates.Open models & compatible serversMore projects & source code
model_idgithub.com1burgerwdev/what-is-jev — Jev notes: fourteen decisions I ran through it · model_id · PythonOpen models & compatible serversMore projects & source code
model_idgithub.com1corasan/arjev — Argent + Jev · model_id · RustOpen models & compatible serversMore projects & source code
oh-my-jevgithub.com0Serve, benchmark, train and compare open System One decision models behind one Jev-compatible gatewayOpen models & compatible serversMore projects & source code
oh-my-jev — READMEgithub.com0Workbench explores and fine-tunes Jev-style decision models, offering a web playground plus accuracy and calibration comparisons against Jev.Open models & compatible serversMore projects & source code
omggithub.com3A System One style decision model runtime in Rust.Open models & compatible serversMore projects & source code
omg — READMEgithub.com3omg is a separate Rust System One-style decision runtime; its README names TypeSafe Jev as the intended contract reference.Open models & compatible serversMore projects & source code
OmniStudio JEV judgmentsgithub.com147Local-first desktop LLM workbench with a JEV typed-judgment page that runs Laya locally via laya-mlx or connects to the cloud, exposing a TypeSafe-compatible endpoint for the official SDKs.Open models & compatible serversMore projects & source code
Open Alternative to Jevgithub.com58Python library that reads option-token probabilities from open-weights models through Transformers or vLLM to give typed, calibrated Choice, Score and Noul decisions in one pass, with temperature scaling and benchmarks.Open models & compatible serversMore projects & source code
open-alternative-jev — READMEgithub.com58A local, open-weight System One package is presented as an alternative to TypeSafe Jev; README performance claims are author-reported.Open models & compatible serversMore projects & source code
open-jev (nico-martin)github.com19Browser-focused TypeScript library that runs open Jev-shaped typed-decision models such as Kev on-device through Transformers.js, on WebGPU or WebAssembly, returning a calibrated distribution per question.Open models & compatible serversMore projects & source code
open-jev demo — projectgithub.com—Browser demo of the open-jev npm package running typed decisions locally with Kev and open-jev ONNX models through Transformers.js, no server or API key.Open models & compatible serversMore projects & source code
Open-Jev — READMEgithub.com381Open-Jev describes an independent Jev-inspired model project that returns typed probabilities from context, questions, and candidates.Open models & compatible serversMore projects & source code
open-jev — READMEgithub.com121OpenJev is a local option-scoring runtime with a documented TypeSafe-style System One endpoint, relevant as a self-hosted alternative; this does not establish tested Jev compatibility.Open models & compatible serversMore projects & source code
open-jev — READMEgithub.com61An unofficial Jev-related research implementation whose README says its weights are random and it is not the production Jev model.Open models & compatible serversMore projects & source code
open-spark-jev — READMEgithub.com21An independent local decision-model implementation inspired by Jev and System One; its README explicitly says it is not TypeSafe Jev or a reproduction of that system.Open models & compatible serversMore projects & source code
OpenDecision — READMEgithub.com57Semantic decision implementation follows Jev’s state-and-typed-question pattern with Choice, Noul and Score primitives.Open models & compatible serversMore projects & source code
openjevgithub.com0TypeSafe Jev ecosystem repository.Open models & compatible serversMore projects & source code
OpenJevgithub.com560Open System One decision server that speaks Jev's wire API, so TypeSafe SDKs work unchanged, reading typed answers off DiffusionGemma 26B-A4B probabilities via vLLM or MLX, including questions about images.Open models & compatible serversMore projects & source code
openjev (daseinlabs)github.com121One-pass option scorer with a local Gemma 3 4B on Apple Silicon via MLX that prefills the context once, scores every option in a single forward pass, and exposes a System One-compatible endpoint.Open models & compatible serversMore projects & source code
OpenJev (GPT-AGI)github.com14Open-source Jev-compatible decision engine that answers Choice, Score and Noul questions from open-weight models in one forward pass behind /v1/systemone, with a Claude Code style REPL and a maze demo.Open models & compatible serversMore projects & source code
openjev — READMEgithub.com3A related open local decision-model implementation with a separate RemoteJev cloud route. Local readout is not TypeSafe Jev, and claims of calibration or zero hallucination are unverified.Open models & compatible serversMore projects & source code
openjev — READMEgithub.com0openjev is a local Gemma 3 4B option scorer that also exposes a TypeSafe-style /v1/systemone request; it is a local alternative, not the hosted TypeSafe Jev model.Open models & compatible serversMore projects & source code
openjev — READMEgithub.com560Razorback OpenJev describes an independent model-backed System One server that returns typed decisions and probabilities.Open models & compatible serversMore projects & source code
openjev-sglang — READMEgithub.com335A self-hosted server implements the TypeSafe/Jev HTTP API using Qwen3.6-35B-A3B on SGLang; the README’s compatibility claim is not independently tested.Open models & compatible serversMore projects & source code
OpenJevProgithub.com18Framework that turns open-weight LLMs such as Qwen3, DeepSeek, Gemma and gpt-oss into typed Choice, Noul and Score decision services using constrained logprob calibration.Open models & compatible serversMore projects & source code
openvonsgithub.com13Open decision layer that makes LLMs, VLMs, and ASR models pick from a finite option set with none-of-the-above, splitting calibrated probabilities into execute, confirm, and reject; includes Japanese voice commands.Open models & compatible serversMore projects & source code
parallelConstraintDecodinggithub.com13Java and Python implementations of parallel constrained decoding that fill a JSON schema of booleans and enums in two forward passes on llama.cpp models, with per-field confidence and a benchmark web app.Open models & compatible serversMore projects & source code
PlayJevgithub.com45Qwen3.5-0.8B fine-tuned as a Jev-like multimodal model that plays ten browser games from raw pixels, one frame in and one move out with a probability per option, in 43 ms on an H200.Open models & compatible serversMore projects & source code
PlayJevgithub.com45🚀🚀 A 0.8B JEV-like multimodal model playing GUI games directly from raw pixels.Open models & compatible serversMore projects & source code
poorjevgithub.com5Local System One-style decision layer that implements Jev's typed Choice, Score and Noul interface on zero-shot NLI models, with temperature scaling and conformal abstention cutting ECE from 0.170 to 0.071.Open models & compatible serversMore projects & source code
primitivesgithub.com3Mintzs/jevify — An optimized inference engine to turn LLMs into Jev-like machines: optimized for quick, lightweight, and accurate decision-making, classification, and scoring · primitives · PythonOpen models & compatible serversMore projects & source code
primitivesgithub.com4,638TheoLeeCJ/SemIf — Semantic ifs from open models, on a 3090 at home. Independent; not affiliated with Jev or TypeSafe. · primitives · PythonOpen models & compatible serversMore projects & source code
primitivesgithub.com12leesk212/JEV-CPU — Run SemIf (Jev-style semantic-if decisions) on a CPU — no GPU. Reads typed option probabilities straight from an open model in one forward pass, plus a web UI. · primitives · PythonOpen models & compatible serversMore projects & source code
ProgramAsWeights — repogithub.com—Research project that compiles an English function description into a LoRA adapter for a frozen Qwen3-0.6B, giving local text classifiers that run on a CPU without an API.Open models & compatible serversMore projects & source code
Qwen Choicegithub.com7Local Jev-style visual judgment on Qwen3.5-4B: an image, a question and options go in, and one forward pass returns the choice with probabilities in about 169 ms on an RTX A6000.Open models & compatible serversMore projects & source code
Qwen3.5-OneForwardgithub.com4Jev-style typed decisions from Qwen3.5-2B logits — one forward pass, zero decoding, zero fine-tuning.Open models & compatible serversMore projects & source code
Qwen3.5-OneForward — READMEgithub.com4An independent research project uses Qwen3.5-2B logits for Jev-inspired typed decisions, explicitly distinguishing itself from TypeSafe Jev.Open models & compatible serversMore projects & source code
Qwen3.8 as Jev-compatible endpointgithub.com—Benchmark recipe serving a stock Qwen3.8-27B W4A16 checkpoint as a no-train Jev-compatible /v1/systemone endpoint on one CMP 170HX card with vLLM, including calibration results.Open models & compatible serversMore projects & source code
Qwen3.8 as Jev-compatible endpoint — repogithub.com—Benchmark recipe serving a stock Qwen3.8-27B W4A16 checkpoint as a no-train Jev-compatible /v1/systemone endpoint on one CMP 170HX card with vLLM, including calibration results.Open models & compatible serversMore projects & source code
ReflexGategithub.com6Offline Go gateway for local agent loops with loop-exit, argument-extraction, and safety-scan endpoints plus a Jev-compatible /v1/systemone for typed noul and choice triage, with a Windows GUI.Open models & compatible serversMore projects & source code
revgithub.com2Typed decisions from a frozen open model, one forward pass; a local JevOpen models & compatible serversMore projects & source code
rev — READMEgithub.com2The README describes Rev as a frozen open model for local typed decisions and says Jev clients can switch to its TypeSafe-shaped endpoint by changing the URL; this review did not test compatibility.Open models & compatible serversMore projects & source code
ruby-layagithub.com24TypeSafe Jev ecosystem repository.Open models & compatible serversMore projects & source code
ruby-laya — READMEgithub.com24ruby-laya ports Laya typed choice, score, and noul decisions over arbitrary state into Ruby.Open models & compatible serversMore projects & source code
ruling — READMEgithub.com6Ruling is an independent local-model alternative that mimics typed, calibrated decisions; its README distinguishes it from TypeSafe Jev while allowing a Jev backend.Open models & compatible serversMore projects & source code
Saik0s/diffusiongemma-jev-macosgithub.com2Local JEV-style decisions with DiffusionGemma on Apple Silicon, with benchmarks and coding-agent examples.Open models & compatible serversMore projects & source code
sarvam-jevgithub.com46Open Jev-style inference engine on Sarvam's Indic sarvam-1 model that reads typed answers from lettered-option logits instead of generating JSON, reusing one state prefill across questions, in the browser.Open models & compatible serversMore projects & source code
sdkgithub.com1ChristianAlexander/effect-jev-cwe — A demonstration of the Jev System 1 model in Effect, matching vulnerabilities to their underlying CWEs · sdk · TypeScriptOpen models & compatible serversMore projects & source code
sdkgithub.com1afanjul/jev-llm — Fake autoregressive language model powered by TypeSafe Jev · sdk · PythonOpen models & compatible serversMore projects & source code
sdkgithub.com7erik-dunteman/ChatJev — Turning the Jev classifier model into an autoregressive next token predictor · sdk · PythonOpen models & compatible serversMore projects & source code
SemIf-OpenJevgithub.com4,639Semantic ifs from open models, on a 3090 at home. Independent; not affiliated with Jev or TypeSafe.Open models & compatible serversMore projects & source code
SemIf-OpenJev — READMEgithub.com4,639SemIf is an independent research project that reads typed option probabilities directly from a model rather than reproducing Jev's undisclosed model.Open models & compatible serversMore projects & source code
simple-jevgithub.com577Reads next-token logits from any Hugging Face model for choice, rubric, and support questions; public demo API with no key.Open models & compatible serversMore projects & source code
snapjudge — READMEgithub.com14Snapjudge runs typed decisions on local Qwen models and says its HTTP API follows TypeSafe System One’s format.Open models & compatible serversMore projects & source code
stuntd — READMEgithub.com51Self-hosted proxy records typed decisions and learns to answer them locally while exposing the Jev System One protocol.Open models & compatible serversMore projects & source code
syn — READMEgithub.com0Qwen3 option scorer offers a Jev-style System One endpoint without generating answers; its probabilities are preferences, not calibrated success rates.Open models & compatible serversMore projects & source code
System Onegithub.com34Minimal code that turns any open LLM into a Jev-style, typesafe-sdk-compatible classifier via batched single-token choice inference; with Qwen3-8B it plays Doom at 172 ms per action versus 600 ms with tool calls.Open models & compatible serversMore projects & source code
System One Litegithub.com41Proof-of-concept HTTP service for constrained decisions with a local language model that answers declared Choice, Score, and Noul questions by reading scores at a fixed answer position, with no generation.Open models & compatible serversMore projects & source code
systemonegithub.com1A Proxy that enables zero-shot system one typed-decisions endpoint for decoder only models served with vLLM or Sglang or Openrouter (Experimental)Open models & compatible serversMore projects & source code
systemone — READMEgithub.com1An independent adapter implements Jev-style typed decisions on other LLM backends, with experimental multimodal support; it is not TypeSafe's model.Open models & compatible serversMore projects & source code
tacetgithub.com2Typed decisions from one encoder pass. Open weights, Apache 2.0.Open models & compatible serversMore projects & source code
tacet — READMEgithub.com2Tacet Sonata is a separate typed-decision model with a documented Choice output; its Jev-format note is a README statement, not verified API compatibility.Open models & compatible serversMore projects & source code
tamev — READMEgithub.com0Independent Jev/Kev alternative scores typed choice, noul and score questions with a lightweight option encoder and exposes a TypeSafe-shaped System One endpoint.Open models & compatible serversMore projects & source code
tensai System One endpointgithub.com—Tiny pure-Go neural-network framework that serves Jev-shaped System One requests from any loaded model by reading labeled logits after each question, so Jev API clients can point at it.Open models & compatible serversMore projects & source code
Tezgithub.com1TypeSafe Jev ecosystem repository.Open models & compatible serversMore projects & source code
Tez — READMEgithub.com1Tez is an open, self-hosted System One alternative inspired by Jev’s published technique, not by Jev outputs.Open models & compatible serversMore projects & source code
truetype.ai-open — READMEgithub.com3Truetype.ai Jev Replica describes a local Gemma 4 12B model as a replica of Jev’s System One API; that is the author’s compatibility claim, not an independent test.Open models & compatible serversMore projects & source code
TurboLLM Jev modegithub.com—Local LLM runner that detects NLI cross-encoder models as local Jev models and serves them through /v1/systemone, /v1/classify and /v1/rerank, with a JSON-first playground.Open models & compatible serversMore projects & source code
typed-lmgithub.com10Single-forward-pass semantic routing in Rust: turn dense LLMs (Llama, Qwen, Mistral, Gemma) into a typed decision API with LoRA/QLoRA training and FP8/FP4 quantization, built on Candle.Open models & compatible serversMore projects & source code
typed-lm — READMEgithub.com10A Rust/Candle implementation that turns decoder models into typed boolean, choice and score decisions, with adapter training and a System One-style HTTP interface. Listed as a related alternative to Jev, not proof of TypeSafe affiliation or drop-in compatibility.Open models & compatible serversMore projects & source code
typedecide — READMEgithub.com0typedecide is an independent browser-based typed-decision readout using Qwen logits, with OpenJev/Sarvam-Jev cited as design references; it is not TypeSafe Jev.Open models & compatible serversMore projects & source code
TypeLLMgithub.com37Jev-inspired SGLang extension that gives open autoregressive LLMs typed outputs (string, integer, number, boolean, enum) from a JSON Schema, with single-token choices, shared-prefix KV reuse and dependency graphs.Open models & compatible serversMore projects & source code
typryxgithub.com0Typed answers with a probability for agent stacks: a choice, a score or a yes/no from a versioned template. Only the fields a template names leave the box, and every answer is on the record. An optional add-on to the TAIPANBOX stack, over HTTP and MCP.Open models & compatible serversMore projects & source code
typryx — READMEgithub.com0Typryx documents Jev as one selectable backend, mapping each request to a typed System One question and converting answers back to its caller-facing form.Open models & compatible serversMore projects & source code
usejev — READMEgithub.com11A local Laya decision-model server using Bun and native ONNX, with a bilingual playground and a TypeSafe-style SDK interface. Its documented model is laya, not the SDK default jev-latest; listed as an independent related alternative, not hosted TypeSafe Jev.Open models & compatible serversMore projects & source code
Valengithub.com604Multimodal decision model on Qwen3.5 that scores candidates against images and video as well as text, with training code and open weights.Open models & compatible serversMore projects & source code
vLLM DiffusionGemma Jev modegithub.com92,393Open vLLM pull request adding a structured generation mode that turns DiffusionGemma into a Jev-like model, with a prototype server for the /v1/systemone endpoint.Open models & compatible serversMore projects & source code
vllm-jev-decisongithub.com8Plugin for vLLM that adds a classification-only mode for Jev-like typed decisions, scoring candidate labels from a finite JSON Schema to return typed values, probabilities, and abstention with no generation.Open models & compatible serversMore projects & source code
webml-kitgithub.com6Framework-agnostic toolkit for running ML models in the browser over WebGPU/WASM that adds a Jev-style decision task, reading typed Choice, Noul and Score answers from local models such as Qwen or OpenJev GGUFs.Open models & compatible serversMore projects & source code
XERJgithub.com—Search engine with optional Jev reranking and a separate local System One-compatible endpoint.Open models & compatible serversMore projects & source code
zwliJay/jev-forgegithub.com105An open training and inference stack for Jev-style decision models. Train models to score dynamic candidate branches from a shared prefix, with support for high-cardinality choice, calibration, and fast batched inference.Open models & compatible serversMore projects & source code
← Back to Awesome Jevgithub.com24Listed by the source without a separate description.Open models & compatible serversMore projects & source code
More posts & discussions44 of 44 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
AESwww.reddit.com—Proof-of-concept decision head on Qwen2.5-1.5B-Instruct that runs one prefill pass and scores candidate answers via an FP32 SwiGLU probe and embedding similarity instead of generating tokens.Open models & compatible serversMore posts & discussions
Bespoke Nimble — demox.com—Open data, recipe, and Qwen3.5-9B LoRA model for Jev-style typed decisions, trained on contrastive negatives made by flipping one fact; reports 90.1% agreement with Jev versus 66.4% for base Qwen.Open models & compatible serversMore posts & discussions
Bespoke Nimble — discussionnews.ycombinator.com—Open data, recipe, and Qwen3.5-9B LoRA model for Jev-style typed decisions, trained on contrastive negatives made by flipping one fact; reports 90.1% agreement with Jev versus 66.4% for base Qwen.Open models & compatible serversMore posts & discussions
Brain DOOM decision policywww.reddit.com—Local Jev-style policy that plays DOOM in real time: a frozen MiniLM encoder reads the observation text and a small trained head picks actions, at 58.3 ms per decision.Open models & compatible serversMore posts & discussions
choosekit — discussionwww.reddit.com—Library that scores a finite set of choices with llama.cpp, Ollama, or OpenRouter models and returns typed decisions with probability distributions, plus an MCP server and a SuperGPQA benchmark against Jev 1.13.Open models & compatible serversMore posts & discussions
cua-s1-forms — demox.com—Tiny Jev-like option scorer (706,048 parameters, 2.8 MB) that rates fill, check, click or skip for every form field in one parallel pass, built as the decision layer for cua-driver.Open models & compatible serversMore posts & discussions
Eugeniusz — demowww.reddit.com—MIT-licensed C++17 library with a C ABI that evaluates text locally and returns a choice, rubric score or truth probability, as an independent alternative to the Jev decision interface for C, C#, Python, Unity and Unreal.Open models & compatible serversMore posts & discussions
Four schools of OpenJEVx.com—Chinese comparison of 8 open Jev-like projects (SemIf, Simple Jev, Laya, Von, Verdict, Kev, Nimble, OpenJev) grouped into four approaches, with reported accuracy and latency.Open models & compatible serversMore posts & discussions
Glance — demox.com—Harness for asking an open vision-language model (Qwen3-VL-4B by default) typed yes/no, pick-one and rating questions about an image, reading calibrated probabilities from one forward pass, with Jev-style request shapes.Open models & compatible serversMore posts & discussions
hearim — demowww.reddit.com—Go gateway that serves the Jev /v1/systemone contract on ordinary LLM backends like Ollama, llama.cpp and vLLM by scoring each question as a single-token choice from logprobs.Open models & compatible serversMore posts & discussions
Jev Quantum — demox.com—Jev-protocol-compatible random baseline in Rust that speaks noul, choice and score but ignores the prompt and answers from a fast pseudo-random generator, for use as a mock or a lower bound in agent-routing evals.Open models & compatible serversMore posts & discussions
Jev Visual — demox.com—Educational Jev-like visual inference on Apple Silicon: Qwen3.5-0.8B on MLX answers choice, yes/no and score questions about one image from shared context and logits, with sorting-factory, Breakout and gesture demos.Open models & compatible serversMore posts & discussions
jevify (fidecastro) — discussionwww.reddit.com—Adapter that serves any model with logprobs behind an OpenAI-compatible endpoint, or a small in-process model, as a Jev-like endpoint by reading typed answers off the next-token distribution in Jev's API shape.Open models & compatible serversMore posts & discussions
jevmlx — discussionnews.ycombinator.com—Local Jev-style decision layer for Apple Silicon that scores every allowed option of boolean, enum and multi-select fields from MLX model logits in one prefill and returns schema-valid JSON with per-field probabilities.Open models & compatible serversMore posts & discussions
kevMac — demowww.reddit.com—Mac app that runs the open Kev decision models locally on Apple Silicon: paste a document, build yes/no, choice or score questions as forms and see calibrated probability bars; kev-0.6b answers in ~0.2 s on an M4.Open models & compatible serversMore posts & discussions
Laya MPS — demonews.ycombinator.com—Runs the open Laya model for Jev-style typed decisions on Apple Silicon via PyTorch MPS, at about 32 ms median latency and 2.1 GiB of RAM on an M5 Pro.Open models & compatible serversMore posts & discussions
Laya MPS — discussionwww.reddit.com—Runs the open Laya model for Jev-style typed decisions on Apple Silicon via PyTorch MPS, at about 32 ms median latency and 2.1 GiB of RAM on an M5 Pro.Open models & compatible serversMore posts & discussions
Laya on Mac M4 CoreMLnews.ycombinator.com—Recipe for running the multilingual Laya decision model offline on an M4 Mac through CoreML on the Neural Engine, with a snake demo, at about 45 decisions per second.Open models & compatible serversMore posts & discussions
Laya vs Jev at Tetrisx.com—Tetris showdown in which the open-weights Laya model, running locally on a 16GB MacBook Air, beats cloud Jev by making decisions 11 times faster.Open models & compatible serversMore posts & discussions
Laya vs Jev Snake battlex.com—Snake duel built with Codex between local Laya and the Jev API on an M5 Pro: median decision time 15.3 ms for Laya versus 298.1 ms for Jev.Open models & compatible serversMore posts & discussions
Laya-MLX vs Jev accuracy testx.com—Two rounds of 100 decision questions on an M2 Pro: local Laya-MLX answered in 13.7ms per question but at about 50% accuracy, while cloud Jev was slower but far more accurate.Open models & compatible serversMore posts & discussions
Laya-MLX — demox.com—Native MLX runtime for the open Laya typed-decision checkpoints on Apple Silicon: 13.4 ms median per short English decision and 7.4 ms with the multilingual checkpoint, with no PyTorch or cloud API; demoed on Snake.Open models & compatible serversMore posts & discussions
Laya-MLX — demo 2x.com—Native MLX runtime for the open Laya typed-decision checkpoints on Apple Silicon: 13.4 ms median per short English decision and 7.4 ms with the multilingual checkpoint, with no PyTorch or cloud API; demoed on Snake.Open models & compatible serversMore posts & discussions
laya.cpp — demowww.reddit.com—Standalone C++ inference for the Jev-style Laya typed-decision models using ggml, with optimized CUDA kernels for NVIDIA RTX GPUs and a Vulkan FP32 backend, covering the english, multilingual and typed-decisions checkpoints.Open models & compatible serversMore posts & discussions
Mevx.com—Decision model with 0.4B parameters for job and candidate matching, fine-tuned from Laya in about 45 minutes; it ties Kev-4B (Spearman 0.733 vs 0.730) but Jev still ranks best at 0.818.Open models & compatible serversMore posts & discussions
mini-Jev — discussionnews.ycombinator.com—Pre-registered experiment showing a Jev-style typed-decision interface on frozen Qwen3-4B by turning schema fields into lettered questions and reading option-letter logits in one pass, compared with grammar-constrained JSON.Open models & compatible serversMore posts & discussions
Open Alternative to Jev — discussionnews.ycombinator.com—Python library that reads option-token probabilities from open-weights models through Transformers or vLLM to give typed, calibrated Choice, Score and Noul decisions in one pass, with temperature scaling and benchmarks.Open models & compatible serversMore posts & discussions
open-jev (nico-martin) — demox.com—Browser-focused TypeScript library that runs open Jev-shaped typed-decision models such as Kev on-device through Transformers.js, on WebGPU or WebAssembly, returning a calibrated distribution per question.Open models & compatible serversMore posts & discussions
openjev (daseinlabs) — demonews.ycombinator.com—One-pass option scorer with a local Gemma 3 4B on Apple Silicon via MLX that prefills the context once, scores every option in a single forward pass, and exposes a System One-compatible endpoint.Open models & compatible serversMore posts & discussions
openjev — demox.com—Qwen3.5 checkpoints trained as a single NLI cross-encoder whose argmax entailment picks the answer, used zero-shot for reranking, grading, guarding and playing Doom from text or pixels.Open models & compatible serversMore posts & discussions
openjev — discussionwww.reddit.com—Qwen3.5 checkpoints trained as a single NLI cross-encoder whose argmax entailment picks the answer, used zero-shot for reranking, grading, guarding and playing Doom from text or pixels.Open models & compatible serversMore posts & discussions
OpenJev — discussionwww.reddit.com—Open System One decision server that speaks Jev's wire API, so TypeSafe SDKs work unchanged, reading typed answers off DiffusionGemma 26B-A4B probabilities via vLLM or MLX, including questions about images.Open models & compatible serversMore posts & discussions
Parallel Constrained Decision Engine — discussionnews.ycombinator.com—Gradio demo of parallel constrained decoding that fills multi-field JSON decision schemas in one pass over Qwen2.5-1.5B on Apple Silicon, pitched as an open alternative to Jev.Open models & compatible serversMore posts & discussions
parallelConstraintDecoding — discussionwww.reddit.com—Java and Python implementations of parallel constrained decoding that fill a JSON schema of booleans and enums in two forward passes on llama.cpp models, with per-field confidence and a benchmark web app.Open models & compatible serversMore posts & discussions
poorjev — discussionwww.reddit.com—Local System One-style decision layer that implements Jev's typed Choice, Score and Noul interface on zero-shot NLI models, with temperature scaling and conformal abstention cutting ECE from 0.170 to 0.071.Open models & compatible serversMore posts & discussions
ProgramAsWeightswww.reddit.com—Research project that compiles an English function description into a LoRA adapter for a frozen Qwen3-0.6B, giving local text classifiers that run on a CPU without an API.Open models & compatible serversMore posts & discussions
Qwen-2.5-1B-RLCD — demo 2x.com—Parallel constrained decoding engine for Apple Silicon that answers multi-field decision schemas in one pass over stock Qwen2.5-1.5B, shipped as code only with no trained RLCD weights.Open models & compatible serversMore posts & discussions
Qwen-2.5-1B-RLCD — postx.com—Parallel constrained decoding engine for Apple Silicon that answers multi-field decision schemas in one pass over stock Qwen2.5-1.5B, shipped as code only with no trained RLCD weights.Open models & compatible serversMore posts & discussions
simple-jev — demox.com—Turns any Hugging Face model into a decision endpoint with a /v1/systemone alias (probabilities are not calibrated).Open models & compatible serversMore posts & discussions
simple-jev — demo 2x.com—Turns any Hugging Face model into a decision endpoint with a /v1/systemone alias (probabilities are not calibrated).Open models & compatible serversMore posts & discussions
simple-jev — demo 3x.com—Turns any Hugging Face model into a decision endpoint with a /v1/systemone alias (probabilities are not calibrated).Open models & compatible serversMore posts & discussions
System One Lite — discussionnews.ycombinator.com—Proof-of-concept HTTP service for constrained decisions with a local language model that answers declared Choice, Score, and Noul questions by reading scores at a fixed answer position, with no generation.Open models & compatible serversMore posts & discussions
VEJI-V2www.reddit.com—MIT-licensed non-autoregressive typed decision model: a 3.3M-parameter head on a frozen multilingual MiniLM that scores options jointly, cites evidence spans, and can abstain.Open models & compatible serversMore posts & discussions
vLLM DiffusionGemma Jev mode — demonews.ycombinator.com—Open vLLM pull request adding a structured generation mode that turns DiffusionGemma into a Jev-like model, with a prototype server for the /v1/systemone endpoint.Open models & compatible serversMore posts & discussions
More videos & channels1 of 1 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
Jev mode for llama.cppwww.youtube.com—Fork of llama.cpp adding a llama-server endpoint that takes instructions, a schema, and state and returns every field with a confidence score; 17.3 ms per decision in bulk on a 2 GB model.Open models & compatible serversMore videos & channels
More packages & releases5 of 5 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
@receptron/laya — npmwww.npmjs.com—Node.js and TypeScript package that runs Laya, the open Jev-compatible decision model by Convai Innovations, through ONNX Runtime and returns choice, score, and noul answers in one forward pass.Open models & compatible serversMore packages & releases
@ruvector/typesafe — npmwww.npmjs.com—Local, network-free typed decisions over sentence embeddings that accept Jev's /v1/systemone request and response shape, returning calibrated confidence and an abstain mass.Open models & compatible serversMore packages & releases
EdgeJev — pypipypi.org—Offline CPU runtime that converts open Jev reimplementations such as laya, kev, NanoJev and PlayJev to ONNX with INT8 quantization behind a protocol-compatible local server; 15.6 ms per question on 4 CPU cores.Open models & compatible serversMore packages & releases
Laya-MLX — pypipypi.org—Native MLX runtime for the open Laya typed-decision checkpoints on Apple Silicon: 13.4 ms median per short English decision and 7.4 ms with the multilingual checkpoint, with no PyTorch or cloud API; demoed on Snake.Open models & compatible serversMore packages & releases
open-jev (nico-martin) — npmwww.npmjs.com—Browser-focused TypeScript library that runs open Jev-shaped typed-decision models such as Kev on-device through Transformers.js, on WebGPU or WebAssembly, returning a calibrated distribution per question.Open models & compatible serversMore packages & releases
More guides & websites100 of 100 matches
Resources, repository stars, descriptions and categories
ResourceStarsDescriptionCategorySave
AES — modelhuggingface.co—Proof-of-concept decision head on Qwen2.5-1.5B-Instruct that runs one prefill pass and scores candidate answers via an FP32 SwiGLU probe and embedding similarity instead of generating tokens.Open models & compatible serversMore guides & websites
AES — projecthuggingface.co—Proof-of-concept decision head on Qwen2.5-1.5B-Instruct that runs one prefill pass and scores candidate answers via an FP32 SwiGLU probe and embedding similarity instead of generating tokens.Open models & compatible serversMore guides & websites
Bespoke Nimble — modelhuggingface.co—Open data, recipe, and Qwen3.5-9B LoRA model for Jev-style typed decisions, trained on contrastive negatives made by flipping one fact; reports 90.1% agreement with Jev versus 66.4% for base Qwen.Open models & compatible serversMore guides & websites
bonzi-1.7b-v1-jevhuggingface.co—Recipe and measurements (no weights) for running Bonsai 1.7B GGUF as a Jev-style typed decision function that returns a probability per option from one forward pass.Open models & compatible serversMore guides & websites
bonzi-27b-v1-jevhuggingface.co—Recipe and measurements (no weights) for running Bonsai 1 27B GGUF as a Jev-style typed decision function that returns a probability per option from one forward pass.Open models & compatible serversMore guides & websites
bonzi-27b-v2-jevhuggingface.co—Recipe and measurements (no weights) for running Ternary Bonsai 2 27B GGUF as a Jev-style typed decision function with one forward pass per decision.Open models & compatible serversMore guides & websites
bonzi-4b-v1-jevhuggingface.co—Recipe and measurements (no weights) for running Bonsai 1 4B GGUF as a Jev-style typed decision function that returns a probability per option from one forward pass.Open models & compatible serversMore guides & websites
bonzi-8b-ternary-v1-jevhuggingface.co—Recipe and measurements (no weights) for running Ternary Bonsai 1 8B GGUF as a Jev-style typed decision function that returns a probability per option from one forward pass.Open models & compatible serversMore guides & websites
bonzi-8b-v1-jevhuggingface.co—Recipe and measurements (no weights) for running Bonsai 1 8B GGUF as a Jev-style typed decision function that returns a probability per option from one forward pass.Open models & compatible serversMore guides & websites
cu-Jev — writeupdtunai.blog—C/CUDA decision engine that turns open Qwen3.5 checkpoints (0.8B to 9B) into a Jev-compatible /v1/systemone API with hand-written kernels and no PyTorch at runtime, plus reproducible benchmarks.Open models & compatible serversMore guides & websites
Decision 1.0huggingface.co—Apache-2.0 collection of open decision models from the vLLM Semantic Router team, 572M encoders and 1.88B to 4.21B decoders that answer Choice, Noul, and Score in one pass across 50 languages, plus a Studio Space.Open models & compatible serversMore guides & websites
Decision 1.0 — modelhuggingface.co—Apache-2.0 collection of open decision models from the vLLM Semantic Router team, 572M encoders and 1.88B to 4.21B decoders that answer Choice, Noul, and Score in one pass across 50 languages, plus a Studio Space.Open models & compatible serversMore guides & websites
Decision 1.0 — spacehuggingface.co—Apache-2.0 collection of open decision models from the vLLM Semantic Router team, 572M encoders and 1.88B to 4.21B decoders that answer Choice, Noul, and Score in one pass across 50 languages, plus a Studio Space.Open models & compatible serversMore guides & websites
ggmlc Laya engine — modelhuggingface.co—Example in the ggmlc neural-network compiler that compiles the open Jev alternative Laya into a standalone C++ decision engine with GGUF weights and a TypeSafe-compatible /v1/systemone server; ~25 ms per decision on an RTX 4050 Laptop.Open models & compatible serversMore guides & websites
Glance — appglance.yohei.me—Harness for asking an open vision-language model (Qwen3-VL-4B by default) typed yes/no, pick-one and rating questions about an image, reading calibrated probabilities from one forward pass, with Jev-style request shapes.Open models & compatible serversMore guides & websites
hedos System One gateway — apphedos.ai—Headless local model engine whose gateway exposes TypeSafe's POST /v1/systemone, so a TypeSafe SDK pointed at it gets answers from a local judge-capable model.Open models & compatible serversMore guides & websites
INSTRUCT_JEV — huggingfacehuggingface.co—Instruction corpus of 119 Choice, Noul and Score rows compiled from TypeSafe's public Jev documentation, published as a HuggingFace dataset.Open models & compatible serversMore guides & websites
Jev mode for llama.cpp — modelhuggingface.co—Fork of llama.cpp adding a llama-server endpoint that takes instructions, a schema, and state and returns every field with a confidence score; 17.3 ms per decision in bulk on a 2 GB model.Open models & compatible serversMore guides & websites
Jev Reproductions Trackerhuggingface.co—Watch: the Jev Reproductions Tracker on Hugging Face follows every open-weight reproduction attempt.Open models & compatible serversMore guides & websites
jev-0.5bhuggingface.co—Prefill-only Jev-style decision model on Qwen2.5-0.5B: a LoRA adapter plus pointer readout head with block-causal attention to answer several questions in one pass behind a /v1/systemone API.Open models & compatible serversMore guides & websites
JEV-CPU — appleesk212.github.io—CPU port of SemIf with a web UI that reads typed option probabilities from Qwen3-0.6B logits in about 1 s, covering eight domains such as moderation, incident severity and credit risk.Open models & compatible serversMore guides & websites
JEV-CPU — modelhuggingface.co—CPU port of SemIf with a web UI that reads typed option probabilities from Qwen3-0.6B logits in about 1 s, covering eight domains such as moderation, incident severity and credit risk.Open models & compatible serversMore guides & websites
jev-gemma-4-E2B-it-choice-64huggingface.co—Browser-oriented ONNX artifact of Gemma 4 E2B with its output projection restricted to 64 fixed answer tokens, turning it into a text-only fixed-label scorer without fine-tuning.Open models & compatible serversMore guides & websites
jev-litehuggingface.co—QLoRA adapter that turns Gemma 4 E4B into a System One decision model, reading answers from the logits at the option letters so it cannot answer outside the given options.Open models & compatible serversMore guides & websites
jev-local-lab decision headshuggingface.co—Tiny trained heads (0.2 to 1.3M parameters) that read structured decisions from a 4-bit Qwen2.5-1.5B on a 24 GB Mac mini, released as a documented side project with honest limits.Open models & compatible serversMore guides & websites
jev-local-lab decision heads — modelhuggingface.co—Tiny trained heads (0.2 to 1.3M parameters) that read structured decisions from a 4-bit Qwen2.5-1.5B on a 24 GB Mac mini, released as a documented side project with honest limits.Open models & compatible serversMore guides & websites
jev-my-brohuggingface.co—Self-hosted English/Thai decision model for agent and tool governance that decides whether to execute, ask for approval, reject or escalate an operation; v0.2 reports 73.78% accuracy.Open models & compatible serversMore guides & websites
jev-my-bro governance datasethuggingface.co—English/Thai dataset of 8,508 cases and 34,032 typed decisions on whether an agent operation should execute, ask the user or be rejected, needs review, is prohibited, and how risky it is.Open models & compatible serversMore guides & websites
Jev-Omnihuggingface.co—Multimodal decision classifier on Gemma 4 12B that returns a probability per option for text, image, audio and video questions; 87.57% on DecisionBench Medium and 86.15% on JevBench.Open models & compatible serversMore guides & websites
jev-on-a-laptop — demohuggingface.co—Unofficial study reproducing Jev-style parallel constrained decoding on stock Qwen 1.5B-8B models on an Apple Silicon laptop, where 7B reached 73.8% agreement against Jev's 86.6% at 0.4-2 s per decision.Open models & compatible serversMore guides & websites
Jev-Style-Qwen3.5-2B-Decision (GGUF)huggingface.co—GGUF builds of a Jev-style Qwen3.5-2B decision model for LM Studio and llama.cpp that reads calibrated option probabilities from one token position; Q4_K_M (1.3 GB) keeps 82.4% accuracy.Open models & compatible serversMore guides & websites
Jev-Style-Qwen3.5-2B-Decision (GGUF) — modelhuggingface.co—GGUF builds of a Jev-style Qwen3.5-2B decision model for LM Studio and llama.cpp that reads calibrated option probabilities from one token position; Q4_K_M (1.3 GB) keeps 82.4% accuracy.Open models & compatible serversMore guides & websites
jev-typed-decisions-causal-0.6bhuggingface.co—LoRA adapters on Qwen3-0.6B-Base trained with the arm-B cached causal typed scorer recipe from pngwn's typed-decisions experiment, a Jev-like reproduction with a frozen LM head.Open models & compatible serversMore guides & websites
jev-typed-decisions-causal-0.6b — apphuggingface.co—LoRA adapters on Qwen3-0.6B-Base trained with the arm-B cached causal typed scorer recipe from pngwn's typed-decisions experiment, a Jev-like reproduction with a frozen LM head.Open models & compatible serversMore guides & websites
JEVfire — sitekikoncuo.github.io—Jev-inspired parallel decisions for CUDA LLMs on vLLM that score single-token labels with the model's own head and assemble JSON in code, with game demos including in-browser Mario at 71 ms per action.Open models & compatible serversMore guides & websites
JevForge — app 2jev-forge.vercel.app—End-to-end stack for Jev-style decision models: synthesizing decision data, training a calibrated Qwen3.5-0.8B candidate scorer, fixed Mind2Web and out-of-distribution evals, preliminary RLCD, and Jev-compatible serving.Open models & compatible serversMore guides & websites
JevForge — modelhuggingface.co—End-to-end stack for Jev-style decision models: synthesizing decision data, training a calibrated Qwen3.5-0.8B candidate scorer, fixed Mind2Web and out-of-distribution evals, preliminary RLCD, and Jev-compatible serving.Open models & compatible serversMore guides & websites
JevForge — model 2huggingface.co—End-to-end stack for Jev-style decision models: synthesizing decision data, training a calibrated Qwen3.5-0.8B candidate scorer, fixed Mind2Web and out-of-distribution evals, preliminary RLCD, and Jev-compatible serving.Open models & compatible serversMore guides & websites
jevify-gemma4-26b-a4bhuggingface.co—Gemma 4 26B-A4B fine-tuned (LoRA, merged) to give honest probabilities on typed questions; the model behind jevify, a local Jev-compatible /v1/systemone decision API.Open models & compatible serversMore guides & websites
jevify-gemma4-26b-a4b GGUFhuggingface.co—Static GGUF quantizations (Q2_K to Q8_0) of jevify-gemma4-26b-a4b, the Gemma 4 model behind the local Jev-compatible jevify decision API.Open models & compatible serversMore guides & websites
jevify-gemma4-e4bhuggingface.co—Gemma 4 E4B fine-tuned (LoRA, merged) to give honest probabilities on typed questions, served by jevify as a local Jev-compatible /v1/systemone decision API.Open models & compatible serversMore guides & websites
jevify-gemma4-e4b GGUFhuggingface.co—Static GGUF quantizations (Q2_K to Q8_0) of jevify-gemma4-e4b, the Gemma 4 E4B model behind the local Jev-compatible jevify decision API.Open models & compatible serversMore guides & websites
jevify-qwen3-vl-2b (Tier 0)huggingface.co—Training-free Jevify recipe that uses Qwen3-VL-2B to answer typed questions about images, reading every allowed answer from one position with a validation-fitted calibration.Open models & compatible serversMore guides & websites
jevify-qwen3-vl-2b-t2huggingface.co—Tier 2 Jevify vision model: Qwen3-VL-2B with a rank-16 decoder LoRA trained on the restricted answer readout with proper scoring rules on A-OKVQA, for typed questions about images.Open models & compatible serversMore guides & websites
jevify-qwen3.5-2bhuggingface.co—Jevify decision heads for Qwen3.5-2B that answer choice, score and noul questions about a state with calibrated probability distributions, without generating text.Open models & compatible serversMore guides & websites
jevify-qwen3.5-4bhuggingface.co—Jevify decision heads for Qwen3.5-4B that read a state, answer choice, score and noul questions and return calibrated distributions, shipping only the heads added on top of the base model.Open models & compatible serversMore guides & websites
jevify-qwen3.5-4b-t2huggingface.co—Tier 2 Jevify model for Qwen3.5-4B that adds a LoRA adapter on top of the decision heads to return calibrated choice, score and noul distributions.Open models & compatible serversMore guides & websites
JevOnehuggingface.co—Typed decision model on Qwen3.6-35B-A3B served through a TypeSafe-compatible /v1/systemone API, with single-token candidate readout, forward and reverse option-order evaluation and calibration.Open models & compatible serversMore guides & websites
jevons-lfm25-1.2b-systemonehuggingface.co—Seed LoRA on LFM2.5-1.2B for jevons, a local System One server that scores allowed outcomes from logits and assembles choice, probabilities, confidence, noul and score in code.Open models & compatible serversMore guides & websites
KaLM-Jev — apphuggingface.co—Local Choice, Score, and Noul judgment service built on KaLM-Reranker-V1-R2 checkpoints in Nano, Small, and Large sizes, returning structured judgments without generating answer text.Open models & compatible serversMore guides & websites
Laya on Mac M4 CoreML — gistgist.github.com—Recipe for running the multilingual Laya decision model offline on an M4 Mac through CoreML on the Neural Engine, with a snake demo, at about 45 decisions per second.Open models & compatible serversMore guides & websites
Laya playground — appbrainfunctioncollapse.com—Local website with games, a benchmark and an agent skill for the open-source Laya decision model, comparing it with hosted Jev on the same 500 labelled examples.Open models & compatible serversMore guides & websites
Laya plays Doommedium.com—Tests Laya, an open 421M-parameter Jev alternative, by having it play Doom locally, adding a dedicated shoot question and logged rule overrides to fix misfires and stuck loops.Open models & compatible serversMore guides & websites
Laya vs Jev at Tetris — appatomic.chat—Tetris showdown in which the open-weights Laya model, running locally on a 16GB MacBook Air, beats cloud Jev by making decisions 11 times faster.Open models & compatible serversMore guides & websites
Laya vs Jev benchmarkhuggingface.co—Independent check of Laya against Jev's published numbers on an RTX 5090: raw Laya sits near chance on PhishNChips phishing (0.505) but reaches 0.611 after Platt calibration, vs Jev's 0.626.Open models & compatible serversMore guides & websites
layaForWebmedium.com—Runs the open 421M-parameter Laya decision model inside a browser tab via ONNX Runtime Web and WebAssembly, shown next to Jev's playground; the browser build matched PyTorch's top answer on all 14 questions.Open models & compatible serversMore guides & websites
layaForWeb — appvishalmysore.github.io—Runs the open 421M-parameter Laya decision model inside a browser tab via ONNX Runtime Web and WebAssembly, shown next to Jev's playground; the browser build matched PyTorch's top answer on all 14 questions.Open models & compatible serversMore guides & websites
LFM2.5-350M-RLCDhuggingface.co—Inference-only parallel structured decoding on unchanged Liquid LFM2.5-350M weights, branching attention and convolution state to score allowed values in a batch, 8-63x faster with field accuracy around 60%.Open models & compatible serversMore guides & websites
LFM2.5-350M-RLCD — relatedhuggingface.co—Inference-only parallel structured decoding on unchanged Liquid LFM2.5-350M weights, branching attention and convolution state to score allowed values in a batch, 8-63x faster with field accuracy around 60%.Open models & compatible serversMore guides & websites
local-system-one-studenthuggingface.co—A 149M ModernBERT encoder distilled from local LLMs that reproduces Jev's interface for one fixed schema, a coding agent's next step, returning tool choice, urgency and destructive-action flags in 19 ms.Open models & compatible serversMore guides & websites
Mev — companymetix.ai—Decision model with 0.4B parameters for job and candidate matching, fine-tuned from Laya in about 45 minutes; it ties Kev-4B (Spearman 0.733 vs 0.730) but Jev still ranks best at 0.818.Open models & compatible serversMore guides & websites
mini-Jev — modelhuggingface.co—Pre-registered experiment showing a Jev-style typed-decision interface on frozen Qwen3-4B by turning schema fields into lettered questions and reading option-letter logits in one pass, compared with grammar-constrained JSON.Open models & compatible serversMore guides & websites
ODM Minihuggingface.co—Open-weight Choice decision head on frozen Qwen3-0.6B that ranks candidate agent actions and returns probabilities, confidence and margin for tool selection (72.97% semantic Choice accuracy).Open models & compatible serversMore guides & websites
OmniStudio JEV judgments — appomnilabs.vibeadmin.cn—Local-first desktop LLM workbench with a JEV typed-judgment page that runs Laya locally via laya-mlx or connects to the cloud, exposing a TypeSafe-compatible endpoint for the official SDKs.Open models & compatible serversMore guides & websites
Open Alternative to Jev — apphuggingface.co—Python library that reads option-token probabilities from open-weights models through Transformers or vLLM to give typed, calibrated Choice, Score and Noul decisions in one pass, with temperature scaling and benchmarks.Open models & compatible serversMore guides & websites
Open Jevhuggingface.co—Hugging Face Space for a Qwen3.5-4B LoRA scorer that answers typed questions with a temperature-scaled probability over the caller's options, encoding the state once and deciding in parallel.Open models & compatible serversMore guides & websites
Open Jev — model 2huggingface.co—Hugging Face Space for a Qwen3.5-4B LoRA scorer that answers typed questions with a temperature-scaled probability over the caller's options, encoding the state once and deciding in parallel.Open models & compatible serversMore guides & websites
Open Jev — model 3huggingface.co—Hugging Face Space for a Qwen3.5-4B LoRA scorer that answers typed questions with a temperature-scaled probability over the caller's options, encoding the state once and deciding in parallel.Open models & compatible serversMore guides & websites
Open Jev — model_v2bhuggingface.co—Hugging Face Space for a Qwen3.5-4B LoRA scorer that answers typed questions with a temperature-scaled probability over the caller's options, encoding the state once and deciding in parallel.Open models & compatible serversMore guides & websites
Open-Jev datasetshuggingface.co—Twelve frozen, mostly synthetic typed-decision data configs (yes/no, choice, multi-label, numeric and ordinal) with manifests, raw records and reconstruction code from the independent Open-Jev project.Open models & compatible serversMore guides & websites
open-jev demohuggingface.co—Browser demo of the open-jev npm package running typed decisions locally with Kev and open-jev ONNX models through Transformers.js, no server or API key.Open models & compatible serversMore guides & websites
open-system-one demohuggingface.co—Demo comparing a 0.1 ms trained decision head with a zero-shot cross-encoder, from an independent benchmark of Jev against open CPU-only alternatives.Open models & compatible serversMore guides & websites
open-system-one results explorerhuggingface.co—Interactive explorer for an independent benchmark of TypeSafe's Jev against open CPU-only stacks on 10,000 decisions.Open models & compatible serversMore guides & websites
open-system-one-benchhuggingface.co—Per-item predictions for 10,000 classification and routing decisions from six stacks, including typesafe/jev and Laya on identical items, so significance tests can be rerun without API spend.Open models & compatible serversMore guides & websites
openjevhuggingface.co—Qwen3.5 checkpoints trained as a single NLI cross-encoder whose argmax entailment picks the answer, used zero-shot for reranking, grading, guarding and playing Doom from text or pixels.Open models & compatible serversMore guides & websites
OpenJev — appcodiv.ai—Open System One decision server that speaks Jev's wire API, so TypeSafe SDKs work unchanged, reading typed answers off DiffusionGemma 26B-A4B probabilities via vLLM or MLX, including questions about images.Open models & compatible serversMore guides & websites
OpenJevPro — siteopenjev.pro—Framework that turns open-weight LLMs such as Qwen3, DeepSeek, Gemma and gpt-oss into typed Choice, Noul and Score decision services using constrained logprob calibration.Open models & compatible serversMore guides & websites
PlayJev — appomnijev.github.io—Qwen3.5-0.8B fine-tuned as a Jev-like multimodal model that plays ten browser games from raw pixels, one frame in and one move out with a probability per option, in 43 ms on an H200.Open models & compatible serversMore guides & websites
PlayJev — modelhuggingface.co—Qwen3.5-0.8B fine-tuned as a Jev-like multimodal model that plays ten browser games from raw pixels, one frame in and one move out with a probability per option, in 43 ms on an H200.Open models & compatible serversMore guides & websites
ProgramAsWeights — modelshuggingface.co—Research project that compiles an English function description into a LoRA adapter for a frozen Qwen3-0.6B, giving local text classifiers that run on a CPU without an API.Open models & compatible serversMore guides & websites
sarvam-jev — appsarvam-jev.feynmanpi.com—Open Jev-style inference engine on Sarvam's Indic sarvam-1 model that reads typed answers from lettered-option logits instead of generating JSON, reusing one state prefill across questions, in the browser.Open models & compatible serversMore guides & websites
Schema-conditioned candidate scorer demohuggingface.co—Interactive demo of a DeBERTa-v3-large scalar-head scorer that reads your own choice, noul and score schema at inference time and decodes grouped logits into typed answers.Open models & compatible serversMore guides & websites
Schema-conditioned candidate scorer demo — modelhuggingface.co—Interactive demo of a DeBERTa-v3-large scalar-head scorer that reads your own choice, noul and score schema at inference time and decodes grouped logits into typed answers.Open models & compatible serversMore guides & websites
simple-jev — sitesimple-jev.featherless.ai—Turns any Hugging Face model into a decision endpoint with a /v1/systemone alias (probabilities are not calibrated).Open models & compatible serversMore guides & websites
solar-mini4-jev — appsolar-mini4-jev.vercel.app—Drop-in wrapper that serves Upstage Solar Mini4 through the Jev System One API shape, with a hosted bring-your-own-key endpoint and a head-to-head benchmark.Open models & compatible serversMore guides & websites
solar-mini4-jev — docshunkim.github.io—Drop-in wrapper that serves Upstage Solar Mini4 through the Jev System One API shape, with a hosted bring-your-own-key endpoint and a head-to-head benchmark.Open models & compatible serversMore guides & websites
Solomonhuggingface.co—LoRA adapter and trained answer heads for Qwen3.8-27B that turn a document plus structured questions into one probability per decision, with pointers to supporting sentences and no text generation.Open models & compatible serversMore guides & websites
System One (gold)huggingface.co—A 70.8M DeBERTa-v3-xsmall cross-encoder for the Jev /v1/systemone schema trained on each dataset's own labels, the gold-supervised arm of a Jev versus Haiku benchmarking study.Open models & compatible serversMore guides & websites
System One (zeroshot)huggingface.co—Unsupervised baseline of a 70.8M DeBERTa-v3-xsmall cross-encoder implementing the Jev /v1/systemone schema, answering Choice, Score and Noul in one batched pass.Open models & compatible serversMore guides & websites
System One Minihuggingface.co—Research prototype with a 69.3M-parameter DistilBERT encoder and five typed classification heads for fixed decisions over synthetic software-diagnosis summaries; explicitly not a Jev reproduction.Open models & compatible serversMore guides & websites
System One vs Laya vs DeepSeekhuggingface.co—Chinese-language Space comparing three decision paths on the same state and typed questions: pngwn's open System One scorer, Laya, and DeepSeek via chat completions returning JSON.Open models & compatible serversMore guides & websites
System One: encode once, decide in parallelhuggingface.co—Space showing the inference path that makes pngwn's System One scorer cheap, encoding the state once and scoring all options in parallel, next to the instruct model generating the same answers.Open models & compatible serversMore guides & websites
System One: encode once, decide in parallel — modelhuggingface.co—Space showing the inference path that makes pngwn's System One scorer cheap, encoding the state once and scoring all options in parallel, next to the instruct model generating the same answers.Open models & compatible serversMore guides & websites
tensai System One endpoint — docsmattn.github.io—Tiny pure-Go neural-network framework that serves Jev-shaped System One requests from any loaded model by reading labeled logits after each question, so Jev API clients can point at it.Open models & compatible serversMore guides & websites
TurboLLM Jev mode — appturbollm.dev—Local LLM runner that detects NLI cross-encoder models as local Jev models and serves them through /v1/systemone, /v1/classify and /v1/rerank, with a JSON-first playground.Open models & compatible serversMore guides & websites
TypeLLM — homepagetypellm.ai—Jev-inspired SGLang extension that gives open autoregressive LLMs typed outputs (string, integer, number, boolean, enum) from a JSON Schema, with single-token choices, shared-prefix KV reuse and dependency graphs.Open models & compatible serversMore guides & websites
VEJI-V2 — modelhuggingface.co—MIT-licensed non-autoregressive typed decision model: a 3.3M-parameter head on a frozen multilingual MiniLM that scores options jointly, cites evidence spans, and can abstain.Open models & compatible serversMore guides & websites
VEJI-V2 — projecthuggingface.co—MIT-licensed non-autoregressive typed decision model: a 3.3M-parameter head on a frozen multilingual MiniLM that scores options jointly, cites evidence spans, and can abstain.Open models & compatible serversMore guides & websites
webml-kit — apphemanth.github.io—Framework-agnostic toolkit for running ML models in the browser over WebGPU/WASM that adds a Jev-style decision task, reading typed Choice, Noul and Score answers from local models such as Qwen or OpenJev GGUFs.Open models & compatible serversMore guides & websites
zwliJay/jev-forge websitejay-forge-web.vercel.app—An open training and inference stack for Jev-style decision models. Train models to score dynamic candidate branches from a shared prefix, with support for high-cardinality choice, calibration, and fast batched inference.Open models & compatible serversMore guides & websites

Showing 569 of 569 resources