Open-source lithos-metal hits 200+ tokens/s/user on Qwen3.8-27B with one M5 Max
JiaZhihao · x · 2026-10-09
lithos-metal is now fully open source: an inference engine that generates Metal megakernels for Apple silicon, combining layer-wise mixer megakernels, GPU-resident generation control, and DSpark speculative decoding. It runs Qwen3.8-27B at 200+ tokens/s/user peak on a single M5 Max, with local serving that plugs into any coding agent in one command. Apache-2.0 licensed, code and tech blog available.
More from coding & agent
- Agents now open the exact app screen where their change lands, cutting eval legwork — lucasmeijer · 2026-10-09
- Dev shows remote workspaces running any agent on any model, incl. vanilla Claude Code — lucasmeijer · 2026-10-09
- Claude Code 2.1.294 fixes instruction-style hooks and premature agent stops — ClaudeCodeLog · 2026-10-09
- Claude Code 2.1.294 fixes instruction-based hooks that failed to block commands and premature agent stops — ClaudeCodeLog · 2026-10-09
- CodeRabbit hits 1,049,817 open-source code reviews in September, 2.49x April's volume — leslysandra · 2026-10-09
- Lucas Meijer open-sources cloud agent coding tool with per-task remote dev environments — lucasmeijer · 2026-10-09