Lithos AI open-sources lithos-metal for local Qwen3.8-27B inference on Apple M5 Max

Lithos AI says it has open-sourced lithos-metal, an inference tool that uses megakernels and DSpark speculative decoding to run Qwen3.8-27B locally.

According to the company's post, the setup reaches a peak of over 200 tokens per second per user on a single Apple M5 Max. The company also says users can try it with any coding agent in one command, with code on GitHub and a technical blog linked. These performance figures are the company's own claims; the report does not include independent benchmarks.