Tag
Introduces cMoLLM, a convolutionally gated mixture-of-LLMs that routes over end-to-end streams through fully differentiable dynamic convolution, improving perplexity and downstream accuracy under matched compute.