Kimi K3 opens up with 2.8T parameters, 1M context, and day-0 vLLM support
青稞AI · wechat · 2026-07-28
A Qwen-hosted talk on Kimi K3 says the model has been open-sourced and got same-day vLLM support.
The session explains how a 2.8T-parameter native multimodal MoE model with 16 active experts out of 896 per token and a 1M-token context stresses inference stacks. It covers the hybrid KDA + periodic full-attention architecture, how vLLM handles cache management and expert parallelism, and a set of optimizations including FlashKDA prefill, fused KDA decode kernels, LatentMoE tail fusion, KV cache offloading, sequence parallelism, and speculative decoding.
Official tests claim DSpark raises single-user generation speed from 118 tok/s to 370 tok/s on SpeedBench, while KDAMetadataBuilder latency drops from 870 µs to 34 µs.
Related event: Kimi K3 shifts attention from scale to architecture(25 posts)→
More from Infra
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11
- LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
- PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11