Jev-like_snake_playground
Local Snake game where code computes facts for each legal move and a local Qwen3.5 picks one in a single Jev-style Choice pass.
Overview
Jev-like Snake Playground is a Chinese-language local Snake game that copies Jev's Choice pattern with a local model instead of the Jev API. Each step, code removes the 180-degree reversal and computes facts for every direction: landing cell, wall or self collision, food eaten, change in distance to food, and reachable empty cells afterward, marking moves with fewer reachable cells than the snake's length as TRAP. Fatal moves are excluded from the options.
With two or more safe moves, Qwen3.5 (0.8B by default) runs one forward pass, scores the single-letter option labels at the last position and applies softmax; with one safe move it is taken directly, and with none the game is declared trapped. A browser page shows the board, option probabilities and a decision log, and runs write summaries and per-step records. It needs an NVIDIA CUDA GPU and does not use TypeSafe or need a key; the method is credited to RJMSWD/QwenJev.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which safe direction should the snake move next? | — |
How it works
- Compute facts for each direction
- Drop fatal moves
- Score remaining options in one Qwen forward pass with softmax
- Move, update the page and log
Strengths
- Fully local, no API key
- Transparent option probabilities
- Code supplies facts so the model only chooses
Caveats
- Requires a CUDA GPU and specific fast kernels
- Not Jev itself; a Jev-like local method
Getting started
Set up a venv with PyTorch for CUDA and the requirements, then `python run.py --local-files-only` and open http://127.0.0.1:8765.At a glance
- Type
- App
- Status
- prototype
- Language
- Python
- Runtime
- Local Qwen3.5 on CUDA with web page
- Has code
- yes
Use cases and integrations
How curated lists describe it
本地贪吃蛇。每步由代码给出合法走法的事实，Qwen 在封闭选项里做一次前向并选出下一步。判断来自 RJMSWD/QwenJev。
hellogumbo/awesome-jev
Frequently asked questions
Does it use Jev?
No, it uses a local Qwen3.5 model with a Jev-style Choice method.
How is a move chosen?
One forward pass scores option letters and a softmax gives probabilities.
What hardware do I need?
An NVIDIA CUDA GPU.
Related projects
Python library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePythonJapanese article testing whether LLMs can replicate Jev, via a logit shortcut on Gemma and the Mario harness.
ArticlediscussionBrowser lab where Jev plays many 2048 games in parallel with one Choice per move and an editor for the state format it receives.
ProjectprototypePythonBrowser arcade games (2048, runner, chess, Flappy Bird, Sudoku) where small decision models choose each move from text options.
ProjectprototypePythonFully local grid tactics game where the open NanoJev model picks each move, attack, heal or dash from real candidate actions.
ProjectresearchPythonKotlin NES emulator whose agent plays Super Mario Bros. and Final Fantasy, with a local model choosing from the goals that apply each turn.
ModelprototypeKotlin