EN Submit a tool
Product

SGLang and Miles Provide Day-One Support for Moonshot AI's 2.8T-Parameter Kimi K3 Model

Published: Source: LMSYS: Blog (Chatbot Arena Team)

ShareXFacebookTelegramWhatsApp

SGLang and Miles provide day-one support for Moonshot AI's open-source 2.8T-parameter model Kimi K3, handling inference and RL training respectively. K3 adopts a hybrid architecture interleaving 69 layers of KDA linear attention with 24 layers of MLA, achieving a single-card batch-1 decoding speed of approximately 113 tok/s on SGLang, and up to about 423 tok/s with DSpark speculative decoding.

Read the original (opens in a new tab)

News stream data aggregated by AI HOT

Related newsLatest in this category
· X: Greg Brockman (@gdb)
· X: Testing Catalog (@testingcatalog)
· X: Kimi.ai (@Kimi_Moonshot)