Laya-MLX Runs Open-Source Decision Model Locally on Mac in Under 1GB, 7.39ms per Decision
量子位 · wechat · 2026-09-22
A week after Jev went viral, developer mizorewww has ported its open-source alternative Laya to Apple MLX. Laya-MLX runs the 421M English model in 943.6MiB peak memory and the 322M multilingual version in just 687.6MiB, fully locally on Apple Silicon with no PyTorch or cloud API.
On an M3 Max, per-decision P50 latency is 7.39ms (322M) and 13.42ms (421M), with throughput up to 395 questions/s on batches of 50. Across 3 checkpoints and two precisions, all 378/378 outputs matched the original model. A Snake demo ran 2,400 steps at 75.4 moves/s with zero deaths.
Laya is a non-autoregressive System 1 decision model (choice/score/boolean) with 10.4k GitHub stars; its typed-decisions accuracy of 76.6% beats Jev's public 72.7%. The trade-off: only 512-1024 token context, best for short, high-frequency structured decisions.
Related event: Open-Source Laya Ported to Apple MLX for 7ms On-Device Decisions(3 posts)→
More from Models
- Dev Builds a Game in an Hour with Tencent's Open-Weights Hy4 (1M Token Context) — mhdfaran · 2026-11-06
- SenseTime launches SenseNova 6.8 Flash with long-term memory, Blender and animated PPT output — heyshrutimishra · 2026-11-06
- Reddit asks: where is MiniMax's promised open-source H3-Regenerate-2K release? — Vyviel · 2026-11-06
- Chinese AI Insiders Say the '3–4 Months Behind' Was Just Distillation Lag — and It's Ending — ronbodkin · 2026-11-06
- Google AI Assistant generates paper summaries — core findings included — from titles alone — EhudReiter · 2026-11-06
- Custom GPTs die Dec 11: if you use someone else's GPT, you can't migrate it — Ok_Negotiation_2587 · 2026-11-06