devwithjevJev-Omni Plays Chrome Dino
Jev-Omni plays Chrome Dino fully locally on an M4 Max, returning probabilities for jump, duck, or nothing based on a cropped slice of road ahead of the dino.
View on X time~100 ms
Erik Kokalj@erik_kokalj๐
Jev-Omni playing Chrome Dino, fully local on an M4 Max ๐ฆ 12B decision model (Gemma 4 finetune) converted to MLX 8-bit, running on M4 Max at ~100 ms. No text generation, it just returns P(jump/duck/nothing) for a slice of road cropped ahead of the dino to cover latency scored 1206 here, best 2412
The 12B decision model is a Gemma 4 finetune converted to MLX 8-bit. It scored 1206 in this run, with a best score of 2412.
Also filed under Games & real time
- Game Agents Powered by Jev
Pitofui integrated Jev into a game and ran it alongside Sonnet Fast.
- Jev Plays DOOM
Jev reads structured state exported by the game and selects targets and behaviors, while a controller executes specific actions.
- Jev Plays The King of Fighters
Jev decides moves for a computer-use agent playing The King of Fighters from Steam on Windows. Sai navigates to start the game and chooses a character, while Jev decides the next move.
- DiffusionGemma-Jev Solves Solitaire
Espen JD built DiffusionGemma-Jev, a Jev-style classifier using DiffusionGemma and vLLM, that solved solitaire on an RTX PRO 6000 in 22 seconds.