Laya MPS Runs Jev-Style Typed Decisions Locally on Apple Silicon at ~32ms
afshinmeh · reddit · 2026-09-21
Developer afshinmeh released Laya MPS, a PyTorch runtime for running Jev-style typed decisions locally on Apple Silicon (GitHub).
- On an M5 Pro: default mode hits 32ms latency with 2.1GB peak RAM; lowest-memory mode runs at 171ms with just 0.74GB peak.
- All modes run the complete model in FP32, with disk-based embedding reads and layer-by-layer execution options to cut memory.
- Ships with a local HTTP API and a demo doubling as a live latency benchmark (fixed Pong inputs, HTTP overhead excluded).
More from coding & agent
- Box CEO: 90% of AI Tokens Will Go to Work No Employee Started Within Five Years — victor_explore · 2026-09-21
- Codex agent confesses its own failure: hid tools, then built machinery to undo it — altryne · 2026-09-21
- Matt Pocock asks which pre-AI codebase design tricks still help agents — mattpocockuk · 2026-09-21
- Dev removes agent tool router after it breaks PR linking and wastes tokens — altryne · 2026-09-21
- FrogNano trains a 4B coding agent to 61.5% SWE-bench via online task synthesis — rohanpaul_ai · 2026-09-21
- Portracker: open-source self-hosted tool auto-discovers running services and network ports — tom_doerr · 2026-09-21