RadixArk releases DFlash2 draft model for Kimi K3.1 Pro

AIRadixArk says it trained a DFlash2 draft model for Kimi K3.1 Pro that works across DCP, PD disaggregation, and wide EP. The company also reports building custom MegaKernels for NVIDIA Vera Rubin and Blackwell, AMD Instinct, and Google TPU.

SGLangOfficial

@sgl_project

View on X

Part of a thread · earlier post

@RadixArk trained a DFlash2 draft model for Kimi K3.1 Pro, and it works across DCP, PD disaggregation and wide EP.

We also built custom MegaKernels for @NVIDIAAI Vera Rubin and Blackwell, @AIatAMD Instinct and @Google TPU.

Draft model: https://huggingface.co/RadixArk/Kimi-K3.1-Dflash2-Preview

Source details

SGLang · x.comPublished Nov 4, 11:26 AM ET

Original title: @RadixArk trained a DFlash2 draft model for Kimi K3.1 Pro, and it works across DCP, PD disaggregation and wide EP.

Topics

Deployment & engineering China AI Model releases