FlashStates v0.1.0

First public release of FlashStates, a serving system that manages the hybrid state of agent sessions: the attention KV and the linear-attention states of hybrid LLMs. It runs in vLLM and in the Mooncake store.

Container images

GPU	Image	Alias	Digest
B200 (x86-64), GLM-5.3-Flash, Qwen3.8-Flash-Next, Nemotron 3 Super	ghcr.io/futuremls-lab/flashstates-b200:v0.33.0-fs3cfd432-mc59ca03f-b200	v0.1.0-b200	sha256:605a03ca0f80ab021063744ea16c40613dbe6b82d066022022811e47dc9aaef3
GB300 (arm64), Kimi-K3.1-Pro	ghcr.io/futuremls-lab/flashstates-gb300:kimi-k31-cu134-266e88c-fs88bad5a-mcf74d68c-gb300	v0.1.0-kimi-k31-gb300	sha256:1191abf4688807e8ba7d60a97b067943e759fc1fabf31113991b83fd957f70ee

The same images are on Docker Hub as weili0234/flashstates (long tags) and weili0234/flashstates-vllm (aliases).

Source

• vLLM: FutureMLS-Lab/vllm, branch flashstate (3cfd432cd8, based on vLLM v0.33.0). The GB300 image is built from branch flashstate-k31pro-gb300 (88bad5a825).
• Mooncake: FutureMLS-Lab/Mooncake, branch flashstate (59ca03f3c4, based on Mooncake v0.3.13.post1). The GB300 image is built from branch flashstate-k31pro-gb300 (f74d68c096).
• Both are pinned as submodules under third_party/. Clone with git clone --recursive.

Links

• Paper: https://www.alphaxiv.org/abs/2611.flashstate
• Project page and demo: https://flashstates.ai
• Getting started, benchmarks and results: see the README.