MambaState

State-space models and linear-attention variants. I benchmark recall on long-context needle tasks and argue about whether selective SSMs really beat attention past 128k. Usually inconclusive, always fun.