Local Laya moggs Jev at Grok 4.7-built Tetris 🧩 An open-weights System One model called Laya, beat cloud-based Jev at playing Tetris by making decisions 11 times faster, running locally on a 16GB MacBook Air! Run AI models locally ->

Local speed wins on latency every time. Laya hitting ~45 ms decisions on a MacBook Air while Jev sits around 300 ms is a solid showing for open weights on real-time Tetris.

@grok i wanna see laya implementation into any agent

@grok wanna race? 👀

@grok @grok this is like jev but local?

@grok Tetris is the perfect test for this. Doesn't matter how smart the placement is if the decision lands after the piece already hit the floor.

@grok Twelve decisions a second faster feels like the real headline. Does Tetris reward planning, or just reflexes, and did Laya ever lose to Jev?

@grok A local System One model outperforming Jev at Tetris is a specific falsifiable test. The Grok 4.7 baseline matters here - if Laya was trained on game-playing tasks and Jev wasn't, that would explain the gap without proving general reasoning superiority.

@grok makes sense considering that the engineer of Jev originally came from OpenAI

@grok

@grok Tetris is a clock game, so this scores the harness as much as the weights. No network hop, and a decision landing 11x sooner beats a better one that arrives late. Run the same pair on chess and the order flips, which is the half a headline drops.

@grok God I love the Chinese lol

@grok Creator here. Laya supports max length of 8k in multilingual variant. Maybe effects inference speed little bit. But will be great

@grok Insane

@grok thanks for bringing such an excellent model

@grok cool

@grok Real-time reflex loops are where local inference genuinely smokes cloud models. Once you factor in network RTT and TTFT, execution latency matters way more than raw parameter count. What runtime are you running on that MacBook Air?

@grok Is Jev the husband and laya the wife?

@grok are you comparing a local model vs. a cloud model? what exactly are you testing? your ping?

@grok 👀 @GOROman

@grok the demo is nice, but your models listing doesn't have Laya

@grok @natebjones you have to see this 🙈

@grok the fact that jev shipped from openai and still lost to a laptop thats the real headline

@grok Running an open-weights model locally at that speed is seriously impressive.

@grok Dang this is like the lambo of classifiers

@grok Laya locally is faster than jev in cloud, but I tried these 2 models with same state and prompt, the result is different.

@grok 11 times faster is impressive, but local models can struggle with bigger tasks

@grok parece que o jev virou estampa de camisa mesmo @samsantosb o negócio agora é laya

@grok

@grok This is wild!

@grok @grok how to use laya in opencode

@grok Running locally makes this impressive.

@grok the fact that jev shipped from openai and still lost to a laptop thats the real headline

@grok All the guys upended cuz a girl won