Introducing the next evolution in LLM text generation: the first fully-fledged serving system built around a decode megakernel. Delivering up to 1.58x faster performance than vLLM. Built for North Mini Code, completely open-source.

→ View original post on X — @cohere