Release assets and suggested copy
These are ready-to-share files and draft text. No social-media post has been submitted.
| Asset | Format | Timing |
|---|---|---|
| Demo video | 1920 × 1080, H.264, 30 video FPS | 30 seconds, original wall-clock speed |
| README GIF | 1040-pixel-wide animated GIF | 15 seconds, original wall-clock speed |
| Maximum-speed video | 1920 × 1080, H.264, 30 video FPS | 15 seconds, original speed, optimized live run |
| Poster | 1920 × 1080 PNG | Actual recorded state at approximately 85 seconds |
| Source recording | JSONL | Entire 100-second actual TTY run |
| Provenance | JSON | Source hash, model, timestamps and renderer hash |
The source run reached score 40 and length 46, with 1,144 real inference-driven moves, zero deaths and zero safety interventions. Its target was 12 decisions/second for legibility. The video is a render of the recorded terminal cells, identified on-screen as RECORDED RUN · 1×.
The extra maximum-speed clip comes from a separate 20.01-second truecolor TTY run with --optimize --max-speed: 1,296 moves at 64.77 moves/second, score 44, length 50, zero deaths and zero interventions. It includes writing the live terminal stream; it is not time-compressed. Source · Provenance.
Short English draft
A local AI that returns probabilities, not generated text.
Laya-MLX runs open-weight typed decision models on Apple Silicon. Watch a 322M model play Snake with a visible cycle safety layer: real probabilities, measured latency, 0 output tokens, no inference API.
One-question API benchmark: 7.39 ms p50 on M3 Max.
pip install laya-mlxCode, weights and reproducible measurements: https://github.com/mizorewww/laya-mlx
中文草稿
让模型直接选方向，而不是先生成一段文字。
Laya-MLX：在 Mac 上本地运行的开放权重决策模型。这个 3.22 亿参数的贪吃蛇 demo，每一步都显示真实方向概率、推理耗时和安全层接管次数。
0 个输出 token，无推理 API。M3 Max 单问题基准 P50 为 7.39 ms。
pip install laya-mlx代码、权重和原始 benchmark：https://github.com/mizorewww/laya-mlx
Separate performance follow-up
The optimized complete Snake loop measured 75.40 moves/second over 2,400 moves, with zero deaths, 2 safety interventions and 2,400/2,400 executed-action agreement with the paired eager control. It was about 6.5% faster in that run. This includes planning, inference, Rich composition, ANSI serialization and game updates, but excludes the terminal emulator’s painting.
Use the optimization report when sharing that number. The 7.39 ms headline describes the separate one-question API fixture; it is not the frame time of this three-question Snake demonstration. Neither result is a cloud-API comparison or evidence of unaided Snake reasoning.