User repository on GitHub
An educational Jev-like visual inference experiment on Apple Silicon: shared context, direct candidate scoring, and local visual demos.
- Forks
- 29
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Jev Visual
An educational Jev-like inference project for Apple Silicon: Qwen3.5-0.8B scores candidate answers from shared multimodal context instead of generating text, with browser, CLI, and HTTP interfaces.
- Stars
- 319
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- hr98w
- Last verified
- Sep 2026
About this project
A small, runnable project for learning vision-language model inference on Apple Silicon. Use Qwen3.5-0.8B with MLX to answer multiple questions about one image: choose an option, judge yes/no, or score ordered levels. Includes a local browser UI, CLI and HTTP API. AI sorting factory Breakout Camera gestures. Classify conveyor objects and choose a sorting lane. Watch video. Locate the ball in five regions to control the paddle. Watch video.
Quoted from the project's README. Read the full README.
Used inJev for Education
Highlights
- Reuses one shared image prefill and forks the cache to score candidate outputs from model logits, avoiding autoregressive structured generation.
- Supports 1 to 64 questions with 2 to 26 options each; probabilities are relative to the supplied options, not correctness estimates.
- Demos cover an AI sorting factory, Breakout, and camera gestures; a recorded Breakout test cleared 9 bricks and returned 6 balls over 80 decisions.
- Needs an Apple Silicon Mac with Metal and downloads about 596 MiB of weights; setup was tested on an M4 with 16 GB running macOS 15.1.
- The README is explicit that it is independent of TypeSafe and does not reproduce Jev's architecture, RLCD training, calibration, or serving.
Quickstart
bash
uv venv --python 3.13
source .venv/bin/activate
uv pip install -r requirements-lock.txt
uv pip install --no-deps -e .
jev-visual-download
JEV_VISUAL_MODEL_PATH=.models/Qwen3.5-0.8B-4bit uvicorn jev_visual.server:app --host 127.0.0.1 --port 8788Watch out
MIT-licensed for the original code with third-party notices. Needs an Apple Silicon Mac with Metal, Python 3.13, uv, and the downloaded Qwen3.5-0.8B-4bit weights; only the Qwen adapter is verified.
Questions about this entry
Short answers drawn from the linked project or page.
A lookalike, by its own admission. The README says it explores a Jev-like inference pattern and explicitly does not claim to reproduce TypeSafe Jev's proprietary architecture, RLCD training, calibration, or serving system. It reuses existing Qwen3.5-0.8B weights and does no training or calibration.
No. The README states candidate probabilities are relative to the supplied options, not correctness estimates. The recorded benchmark of 37.30s down to 2.40s at 64 decisions is a scaling experiment against generated JSON, explicitly not an accuracy evaluation and not a Jev comparison.
Not documented as supported. Setup requires an Apple Silicon Mac with Metal, and it was tested on an M4 with 16GB running macOS 15.1 and Python 3.13, with about 596 MiB of Qwen3.5-0.8B-4bit weights downloaded on first setup. Only the Qwen3.5 adapter is verified.
Barely, and the README concedes as much. Breakout is reduced to asking which of five numbered regions holds the ball, with the paddle moving to a fixed region center; a recorded 80-decision test cleared 9 bricks with 6 returns and 2 lives and stopped before completion. Those trials do not isolate the effect of 4-bit quantization.
Requires an Apple Silicon Mac with Metal, tested on M4 with 16GB, macOS 15.1 and Python 3.13.1. Use `uv venv --python 3.13`, install requirements-lock.txt, run `jev-visual-download` (about 596 MiB of weights), then start uvicorn on 127.0.0.1:8788 with JEV_VISUAL_MODEL_PATH set.
It supports 1 to 64 questions with 2 to 26 options each. Candidate probabilities are relative to the supplied options, not correctness estimates, and only the Qwen3.5 adapter is verified. It uses existing weights with no training or calibration.
It compares generate JSON, independent candidate scoring and shared-prefix batched scoring at 1, 4, 16 and 64 decisions with three repetitions and model loading excluded. The recorded M4/16GB run took 37.30s to 2.40s for independent versus shared scoring at 64 decisions, medians.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
An Apple Silicon decision layer that scores every allowed option of a schema (booleans, enums, multi-selects) from MLX model logits in one prefill, assembles schema-valid JSON itself, and exposes a System One-compatible endpoint.
Jev Repos & SDKs#community#python#open-models
Community
bnsd55/jevmlxPython
bnsd55/jevmlxPythonAn open 0.6B replica of Jev that turns states and questions into full probability distributions without decoding answer tokens, trained and evaluated on Maze, Snake, and ViZDoom.
Jev Repos & SDKs#community#python#open-models
Community
TianyuCodings/NanoJevPython
TianyuCodings/NanoJevPythonAn open, non-autoregressive System One model with published weights and a TypeSafe-compatible /v1/systemone API; the 395M-parameter OptionMarker checkpoint scores discrete and ordinal questions locally.
Jev Repos & SDKs#community#python#open-models
Community
wfzyx/vonPython
wfzyx/vonPythonTurns local text and vision models into Jev-style decision models by reading probabilities from prefill logits alone, with SGLang, Transformers, and Apple Silicon MLX backends.
Jev Repos & SDKs#community#python#open-models
Community
Yinsongxu/LLM2JevPython
Yinsongxu/LLM2JevPythonAn open 0.6B System One decision model that reads unstructured state and typed questions in one forward pass, returning calibrated distributions with zero decoded tokens; weights are published on Hugging Face.
Jev Repos & SDKs#community#python#open-models
Community
malevrigns/agent-jevPython
malevrigns/agent-jevPythonA reproduction of Jev that turns any Qwen checkpoint into a decision model serving the same /v1/systemone schema (Choice, Score, Noul), with no training and no generated answer text.
Jev Repos & SDKs#community#python#open-models
Community
zhengxuyu/litjevPython
zhengxuyu/litjevPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.