Laya runs at 84ms but loses every game: fast local models aren't smarter
eter_inquirer · x · 2026-09-22
tdinhme benchmarked local models on a MacBook via chess and Tetris: Laya-mlx is blazing fast (84ms) but notably dumber, losing 3-0 to Jev. eterinquirer saw the same — Laya is quick but loses every chess match. Takeaway: latency gains in local small models may come at a steep capability cost, and games make a handy sanity check.
More from Models
- Unverified DataBench Charts Fuel Rumors of OpenAI's Internal Model 'Luna' Ahead of GPT-6 — almmaasoglu · 2026-09-22
- OpenAI's 24-day-old internal model reportedly solved 100+ open math problems — IgorCarron · 2026-09-22
- Paradigm unveils Limite 1B 'Violetto', a 1B model built for high-frequency mathematical intelligence — tensorqt · 2026-09-22
- Speculation: Grok Pro line is an extension of Flash line, mxfp4 QAT likely speeds RL rollouts — stochasticchasm · 2026-09-22
- Adam-to-Muon mid-training switch sparks debate, seemingly contradicting Moonlight paper — stochasticchasm · 2026-09-22
- Diffusion LMs were 5-10x faster but never loved — timing, not speed, wins — eliza_luth · 2026-09-22