Kimi K3.1 mixes DSA with recurrent Kimi Delta Attention in one packed cache
vllm_project · x · 2026-11-05
The vLLM project describes Kimi K3.1's architecture: DSA sparse attention combined with recurrent Kimi Delta Attention. A single packed cache holds MLA latents, indexer keys and KDA state; the serving work adds token-sharded prefill, exact top-k merging for decode context parallelism, and a fused low-latency decode path.
Related event: vLLM Ships Day-0 Support for Kimi K3.1 Across Training and Serving(5 posts)→
More from Infra
- Bittensor's Targon hits record $1.2M October revenue, first $1M month — markjeffrey · 2026-11-05
- Frontier training without megacampuses: MoE model trained across <20MW clusters over 25Gbps — _arohan_ · 2026-11-05
- LFM3-Omni-2B ported to run fully offline on iPhone at ~80 tok/s via custom WebGPU kernels — victormustar · 2026-11-05
- Solo dev's open-source browser agent uses local models to watch your screen and trigger actions — Roy3838 · 2026-11-05
- Mistral Large 4 1T hits HF Inference Endpoints at $10/hr on 4x A100 — ngxson · 2026-11-05
- NVIDIA accelerates structural simulation in Abaqus with CUDA-X — NVIDIAAI · 2026-11-05