SGLang says Kimi K3 hits 423 tok/s on day 0 with fused KDA kernels and prefix caching
ying11231 · x · 2026-07-29
- SGLang reports 423 tok/s on Kimi K3 at day 0, measured on GSM8K, and says RL support is already ready in Miles.
- The team says K3’s new architecture was natively implemented and deeply optimized with fused KDA decode kernels, DP attention, DSpark, PD disagg, and KDA-aware prefix caching.
- They also claim to have passed Kimi Vendor Verifier and say the stack is ready for production.
- The post credits collaboration with Kimi Moonshot, NVIDIA, AMD, KVCache AI, Modal, Baseten, and several serving partners.
- Blog, cookbook, and benchmarks were promised in the comments, and the demo video was described as being generated by Kimi K3 itself.
Related event: SGLang Day-0 Support for Kimi K3 Boosts Throughput to 423 tok/s(8 posts)→
More from Infra
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- Wan2GP AMD one-click installer lands on Pinokio, runs AI video on 6GB VRAM — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27