Skip to content
View original post on X: SGLangOfficial
Published
Pick62/100

SGLang announces day-zero support for Kimi K3.1 Pro with sparse attention

AISGLang says it supports Kimi K3.1 Pro on day zero, after the model replaces dense MLA with sparse attention (DSA). It reports up to 33.7K tokens per second per GPU on AgentX with GB300 and 450 tokens per second for a single user with DFlash2 and MegaKernel.

AIWhy it matters.

The post names the architecture change and the specific throughput and hardware support, so readers can judge how serving the model differs from the dense MLA version.

SGLangVerified on XOfficial
@sgl_project
View on X

Kimi K3.1 Pro replaces dense MLA with sparse attention (DSA). It's much more efficient and easier to serve.

SGLang delivers on Day 0:
· Up to 33.7K tok/s per GPU (AgentX, GB300)
· 450 tok/s single user with DFlash2 + MegaKernel
· NVIDIA Vera Rubin / Blackwell / Hopper, AMD, TPU, Trainium, PPU

Kimi.ai@Kimi_Moonshot
Releasing the model weights of Kimi K3.1 Pro: https://huggingface.co/moonshotai/Kimi-K3.1-Pro We'll keep advancing open intelligence.
View quoted post on X
Source details

SGLang · x.comPublished

Original title: Kimi K3.1 Pro replaces dense MLA with sparse attention (DSA). It's much more efficient and easier to serve.

Topics

Open-source ecosystemDeployment & engineeringChina AIProduct updates