jev-visual
Educational MLX project scoring typed questions about one image with Qwen3.5-0.8B on Apple Silicon, with game demos.
Overview
Jev Visual is a small, runnable project for learning vision-language inference on Apple Silicon. It uses Qwen3.5-0.8B with MLX to answer several questions about one image (choose an option, judge yes or no, or score ordered levels) by reusing the shared multimodal context and scoring candidate outputs directly from logits, avoiding autoregressive structured generation. It includes a local browser UI, a CLI and an HTTP API.
Three demos show the pattern: an AI sorting factory that classifies conveyor objects and chooses a lane, Breakout where the model locates the ball in one of five regions to move the paddle, and camera gesture recognition controlling particles. The README is candid that Breakout exposed the small 4-bit model's limits: following the ball or choosing left/right was unreliable until the task was simplified. It states plainly that it is an independent community project that does not reproduce TypeSafe Jev's architecture, RLCD training, calibration or serving, and listings note its probabilities are not calibrated correctness estimates. Documentation is in English and Chinese.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which lane should this conveyor object be sorted into? | — |
| Choice | Which of five regions contains the ball? | — |
| Choice | Which hand gesture is shown in this camera frame?(inferred) | — |
How it works
- Encode the image once as shared context
- Score candidate answers for each question from logits
- Return choice, yes/no or score per question
- Demos act on the answers (sort, move paddle, control particles)
Strengths
- Runs locally on Apple Silicon
- One pass of shared context for several questions
- Honest about model limits and non-calibrated probabilities
Caveats
- Educational; not a Jev reproduction
- Probabilities are not calibrated
- Small model struggles with fine control tasks
At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- MLX on Apple Silicon
- Has code
- yes
- License
- MIT
- Stars
- 317
Use cases and integrations
How curated lists describe it
Educational Jev-like visual inference on Apple Silicon: shared multimodal context, candidate scoring, sorting-factory / Breakout / gesture demos
AnotiaWang/awesome-decision-modelsLocal Jev-like visual inference experiment on Apple Silicon Mac. Scores and classifies single images across multiple questions with 3 playable game demos.
logicrw/awesome-jev-projectsAn educational Jev-like visual inference experiment on Apple Silicon: shared context, direct candidate scoring, and local visual demos.
wh000wh000/awesome-jev-liveEducational MLX/Qwen vision-language experiment sharing image context across candidate-scoring questions; its probabilities are not calibrated correctness estimates.
cobanov/awesome-jevAn educational Jev-like visual inference experiment on Apple Silicon: shared context, direct candidate scoring, and local visual demos.
hellogumbo/awesome-jevMultiple typed questions about one image in a single pass, on Qwen3.5-0.8B with MLX on Apple Silicon. States plainly that it explores the pattern and does not claim to reproduce Jev's architecture or training. Chinese and English.
MrJev/awesome-jev
Frequently asked questions
Is this TypeSafe's Jev?
No. It is an independent educational project exploring a Jev-like inference pattern on Qwen3.5-0.8B.
Are the probabilities calibrated?
No, they are not calibrated correctness estimates.
What hardware does it need?
An Apple Silicon Mac running MLX.
Related projects
StartLux's open typed decision models from 0.8B to 27B plus a 35B-A3B MoE, reading text, JSON or images up to 256K tokens.
ModelresearchPythonPlayJev: Qwen3.5-0.8B fine-tuned to play ten browser games from raw pixels, one forward pass per move with a probability per option.
ServiceresearchPythonFree hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPythonPython library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPython