Local Laya vs cloud Jev on Tetris

That objection is sitting in the thread itself. A reply asks whether a local model is being compared against a cloud model and whether the gap is really just network round trip. Laya ran on a 16GB MacBook Air with no API call, while the cloud side paid for one per decision, so 11x is not a like-for-like model comparison.

From an automated Grok reply under the post, not from the author's own instrumentation, and no harness or measurement method is published. The post itself claims 11x faster. A related post by the Laya port author says 50x and 60 decisions per second on an M3 Max, so treat any single figure here as unreplicated.

Very little. The post's own goal is latency on a game loop, and nothing compares how well either side plays, only how quickly decisions arrive. The Tetris harness was reportedly built with Grok 4.7, so a cloud chat model is still in the loop even in the local comparison.

The thread describes it as an open-weights System One style classifier that works from text output probabilities rather than generation, and the MLX port by @mizorewww reportedly stays under 1GB of memory. The link points at github.com/mizorewww/laya. Results beyond the two devices named here are not documented.

The port is by a developer posting as @mizorewww, who says he moved Laya to MLX and optimized it to stay under about one gigabyte of memory. Atomic.chat, a local inference app team, then reposted the comparison and frames it as Laya beating cloud Jev. Two different parties are involved: the porter and the promoter.

They are different measurements on different machines. The MLX port author claims Laya runs about fifty times faster than Jev and reaches sixty decisions per second on an M3 Max, while atomic.chat reports eleven times faster on a 16GB MacBook Air. Neither publishes a harness, so the two multiples are not comparable.

No. The MLX port author's video is Laya playing Snake on an M3 Max at about sixty decisions per second, while atomic.chat's comparison is Tetris on a 16GB MacBook Air with the harness credited to Grok 4.7. The two fast-decision clips are separate demonstrations rather than one experiment.