{"name":"FlashStates","repo":"FutureMLS-Lab/FlashStates","url":"https://midorreal.com/project/futuremls-lab-flashstates","repository":"https://github.com/FutureMLS-Lab/FlashStates","description":"","facts":{"license":{"value":"Apache-2.0","source":"https://github.com/FutureMLS-Lab/FlashStates","checked":"22 Nov 2026"},"language":{"value":"Python","source":"https://github.com/FutureMLS-Lab/FlashStates","checked":"22 Nov 2026"},"last_release":{"value":"v0.1.0, 9 Nov 2026","source":"https://github.com/FutureMLS-Lab/FlashStates/releases","checked":"22 Nov 2026"},"releases":{"value":"1","source":"https://github.com/FutureMLS-Lab/FlashStates/releases","checked":"22 Nov 2026"},"contributors":{"value":"2","source":"https://github.com/FutureMLS-Lab/FlashStates/graphs/contributors","checked":"22 Nov 2026"},"archived":{"value":"no","source":"https://github.com/FutureMLS-Lab/FlashStates","checked":"22 Nov 2026"}},"stars":286.0,"writeup":{"what_it_is":"FlashStates is a research codebase for managing hybrid model state (attention KV cache plus recurrent checkpoints) in LLM inference for agentic, multi-turn workloads. It includes a modified vLLM v0.33.0 and a modified Mooncake store, along with Docker images, benchmark scripts and results.","audience":"Researchers and engineers running hybrid LLMs for coding-agent workloads","claims":[{"kind":"specific","claim":"Saves recurrent checkpoints at any token position, so the next turn can resume at the exact end of the previous turn","excerpt":"Recurrent checkpoints are saved at any token position","status":"not_checked"},{"kind":"specific","claim":"Reaches up to 3.02× vLLM's decode throughput on closed-loop coding-agent replays with two B200 GPUs per server","excerpt":"FlashStates reaches up to **3.02×**","status":"not_checked"},{"kind":"specific","claim":"Cuts median uncached prompt tokens per request by 96% on GLM-5.3-Flash at 48 concurrent sessions","excerpt":"cuts the median uncached prompt tokens per request","status":"not_checked"},{"kind":"specific","claim":"Provides Docker images for B200 and GB300 GPUs","excerpt":"A B200 image for GLM-5.3-Flash NVFP4","status":"not_checked"},{"kind":"broad","claim":"Provides a cost-aware policy for admitting and evicting checkpoints and KV blocks together","excerpt":"A recurrent checkpoint and the KV blocks it needs are admitted and evicted","status":"too_broad"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-14"},"why_now":null,"owner_supplied":null,"score":{"verdict":"overhyped","hype":61,"reality":30,"gap":31,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T07:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-09T14:01:54+00:00","text":"Latest release v0.1.0","source":"https://github.com/FutureMLS-Lab/FlashStates/releases"},{"date":"2026-11-09T12:30:24+00:00","text":"Repository created","source":"https://github.com/FutureMLS-Lab/FlashStates"}],"cite":"FlashStates is a GitHub project at https://github.com/FutureMLS-Lab/FlashStates written mainly in Python.","attribution":"Data from Mid or Real, https://midorreal.com/project/futuremls-lab-flashstates, checked 22 November 2026. Attribution with a link is required.","still_real":null}