Nanojev Arena

A code planner first excludes reversing, wall, and self collisions, then BFS finds the static shortest path to food and narrows to 2-4 candidates. The model answers one choice question per batch with argmax, while the planner covers until the response arrives.

Run install.py to clone upstream NanoJev, create the virtual environment, install dependencies, and download the roughly 2.3 GB checkpoint over the configured mirror, then start serve_decisions.py and open snake.html on port 8766.

No model output is needed for those: arrows or WASD steer, P pauses, R restarts, wall or body contact ends the run, bodies become food, and the scoreboard ranks live while the side panel shows each AI snake distribution.

Without CUDA you can run the upstream browser replays such as side-by-side pages, which need no model. On Turing cards pass fp32 precision, and CPU-only PyTorch installs should be replaced with a CUDA build.

The measured claim is that the model picks well when it is allowed to pick: the question format follows the upstream training data strictly, using the composed protocol, and the target direction comes back with 97.5% overwhelming probability. Over a full grid NanoJev ate 27 food and survived to the step limit, while the original Qwen3-0.6B got stuck at step 211. The per-frame planner is what stops the snakes driving into a wall while they wait.

It is a 0.6B open local replica of Jev, TypeSafe's System One decision model, from the TianyuCodings/NanoJev project. install.py clones that upstream repository, creates the virtual environment, installs dependencies, downloads the trained checkpoint and copies web/snake.html into the upstream web directory. The weights themselves are published as C-Tianyu/NanoJev on Hugging Face.

This repository's code is MIT. Upstream NanoJev and the checkpoint weights stay with their authors and are likewise MIT and publicly downloadable. Attribution goes to TianyuCodings/NanoJev for the model, training pipeline and server, and to Jev and TypeSafe for the original System One concept. The stated platform support is Windows and Linux.