Announcement
lithos-metal is being open-sourced, with a claimed peak of 200-plus tokens per second per user on one Apple M5 Max
The developer says megakernels and DSpark speculative decoding run Qwen3.8-27B at that peak, and shares the code.
TLDR
The lithos-metal developer says it is open-sourcing the code and claims megakernels plus DSpark speculative decoding run Qwen3.8-27B at a peak of 200-plus tokens per second per user on one Apple M5 Max. The developer also says it can be tried with any coding agent in one command.
Combined views
11.4K
3 Sources, first seen ago
