Picture published by the creator on GitHub.
Watch a tiny open model out-play Jev at Doom, maze, and Snake
Made by Tianyu
View profile on GitHub (opens in a new tab)NanoJev is a small (0.6-billion-parameter) open reproduction of Jev: a model that reads a game or task state and a set of possible actions and returns a full probability distribution over them in one pass, with no text generation.
Its creator trained one shared checkpoint to play four tasks — a ViZDoom aiming game, a moving-target shooting variant, a 50×50 maze, and Snake — and reports it beating the original Jev head-to-head: 128 of 128 successes on the basic Doom task against 56 of 128 for Jev, as of a September 20 update. You can watch it play in a free browser demo with nothing installed, or download the model and training data from Hugging Face.
You could use it to…
- Watch a tiny 0.6B model aim and fire in Doom in real time
- Compare it side by side against Jev and an untrained baseline model
- See the same checkpoint solve a maze and play Snake with no retraining
Can I use this?
You'll need A web browser for the live demo; Python and a GPU only if you want to train or run it yourself · No install needed to watch it play
Worth knowing This 'use' line describes the free hosted playground linked from the README, not the training/inference code itself, which needs Python and a GPU. The head-to-head numbers against Jev are the creator's own benchmark on four specific game tasks. BuildTube has not run or verified this software.
Why it's here
Added 25 September
An open 0.6B model built the same way as Jev, shown beating it live at Doom and Snake.
Part of “The Jev wave”, a BuildTube editorial story
New Sept 17 · 2,239 GitHub stars in about a week
- 2,239 people have starred it on GitHub.
Numbers checked on 25 September 2026; not refreshed since.
Behind it
See the code on GitHubBuildTube has not run or verified this project. Everything above is written from what the creator published.
Made this?