Tetris Bench: Jev vs Laya-MLX

Still from Tetris Bench: Jev vs Laya-MLX — Still from x.com

Post on X

Views: 72,879
Likes: 398
Reposts: 22

Read from X on 24 Sep 2026. Counts change daily.

Reported by the author

Laya latency: ~84 msIt's true that the model is really fast (~84ms)x.com
Result: Jev won 3/3Losing to Jev 3 out of 3 rounds.x.com

Tony Dinh posted a short video of a Tetris match between Jev and Laya, an open-source classifier built along the same lines as Jev. He ran Laya on his MacBook through laya-mlx, mizorewww’s port of the model to Apple’s MLX framework. He reports that Laya is fast, about 84 ms per decision, but “much more stupid”, and that it lost to Jev in all three rounds.

His post quotes mizorewww’s announcement of the port. That post says laya-mlx uses at most 1 GB of memory and plays Snake at 60 decisions per second on an M3 Max. The port is at github.com/mizorewww/laya-mlx.

Open on X

View the post on X

PrimitivesChoice

Use casesReal-time control

Jev Ultrafast (Browser Use)

Browser Use's fast browser agent where Jev picks the operation and the DOM element in a single request; a small LLM is only invoked to write text when typing is required. Demoed on a Google Flights search.

browser-useOpen

19k starsFlight search 7.1 s

ProjectTyped tool dispatch+2

Laya-MLX

An independent MLX port of Laya for Apple Silicon, 13.4ms median for a short English decision and 7.4ms on the multilingual checkpoint, with a Snake demo that plays at up to 75 moves per second under 1GB of memory.

mizorewwwOpen

6.1k starsShort decision 13.4 ms P50Snake demo 75.4 moves/s

ToolReal-time control

Agent Desktop

Rust desktop automation that reads apps through the OS accessibility tree instead of pixels. Its jev-desktop skill takes a plain-language goal and has Jev pick one operation and one target element per turn, so the 150-element tree never enters the agent's context.

lahfirOpen

1.6k starsToken reduction 78-96%

ProjectTyped tool dispatch+1

Laya-CoreML

Core ML ports of all three Laya checkpoints for Apple Silicon, reaching 4.98ms P50 on the Neural Engine for a short multilingual decision, with a Snake demo sustaining about 50 decisions per second.