User repository on GitHub
Turning an LLM model into a Jev like System.
- Forks
- 10
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
OpenSourceJev
Local System One decision engine that turns open-weight Qwen models into typed decision endpoints via direct llama.cpp logits, with fast and accuracy profiles for consumer GPUs.
- Stars
- 42
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- sabeel111
- Last verified
- Oct 2026
About this project
OpenSourceJev is an open-source research experiment exploring a local implementation of the "System One" AI decision paradigm popularized by TypeSafe AI's Jev. Instead of using Large Language Models (LLMs) to generate verbose text or stream fragile JSON blobs that require retry loops and regex parsing, OpenSourceJev turns open-weight models (like Qwen3-1.7B) into ultra-fast, deterministic, strictly-typed decision engines.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Dual fast and accuracy profiles trade latency and VRAM against benchmark and routing accuracy.
- Four native primitives cover yes/no, categorical choice, ordered score expectation, and short text.
- Temperature calibration, SHA-256 provenance, token counts, and runtime parameters ship per decision.
- Bundles a web playground, conditional workflows, and a realtime ViZDoom agent demo.
Watch out
Explicitly an independent research experiment, not a production service and not affiliated with or endorsed by TypeSafe AI. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
It reads raw float32 logits through ctypes straight from the native shared libraries, bypassing chat APIs, then projects over the candidate set with candidate-only softmax. Multi-token phrases accumulate conditional log-probabilities with length normalization.
The fast profile uses a smaller quantized model with lower VRAM and latency, while the accuracy profile uses a larger quantized model with higher benchmark scores at higher latency and VRAM. Both fit the documented 4GB-GPU envelope with pool eviction before loading.
Consumer-laptop class hardware with 8GB RAM and Python 3.11+, CUDA optional since CPU inference is supported. You supply GGUF weights into the models directory or point at an external path; Windows uses a setup script and Linux or macOS needs OpenMP libraries.
Yes: a ViZDoom controller queries the engine in realtime to pick aim, strafe, and shoot actions from game state, alongside a web playground with traces and a conditional workflow engine. Swagger docs expose the HTTP API.
Three ways: pass "profile": "accuracy" or "fast" in the POST /v1/systemone or /api/run payload, set the JEV_PROFILE environment variable, or inspect GET /v1/models for checksums and default parameters. Fast is the default.
Noul returns a calibrated probability for a yes/no condition, Choice selects from a fixed option set with confidence, and Score returns a continuous expectation across ordered rubric levels. Text optionally returns a short generated action or string continuation when one is needed.
Yes. Every decision returns full SHA-256 model provenance, the temperature used, context token counts, and runtime parameters, so a result ties back to the exact weights and settings. Model-scoped temperature is calibrated against Google BoolQ held-out splits, roughly 9.47 for the 1.7B and 1.24 for the 3.5B.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython45GitHub stars
Drop-in service that exposes Upstage Solar models through the System One shape, accepting state plus noul, choice, or score questions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
hunkim/solar-mini4-jevPython
hunkim/solar-mini4-jevPython38GitHub stars
Research recipe that fuses two frozen open readers into Jev-style yes/no judgments with a fit-free router, confidence gate, and candidate sets.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
FeiLiuEM/open-medical-jevPython
FeiLiuEM/open-medical-jevPython62GitHub stars
Simulation study that drives one xArm7 apple-to-plate task with Jev, a large reasoning model, and a small model through identical incremental controls.
Jev Demos & Experiments#community#jevbest#catalog-import
40GitHub stars
Open research toolkit that encodes an image once and answers multiple typed choice, noul, and score questions from a shared distribution, plus a local 0.6B text-decision model.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
IamBusy/OpenJev-VisionPython
IamBusy/OpenJev-VisionPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.