vLLM Ships Day-0 Support for Kimi K3.1 Across Training and Serving

The vLLM project announced full Day-0 support for Kimi K3.1 on the day of its release, covering the entire pipeline from inference serving to reinforcement learning training—the core of this event. The team published multiple posts introducing K3.1 from ecosystem, architecture, and performance angles.

Confirmed

vLLM partnered with inferact on a vLLM + Megatron solution providing Day-0 RL support for Kimi K3.1, covering train-inference consistency for Kimi Delta Attention and supporting long-horizon agent serving scenarios
Hardware support is broad: NVIDIA GPUs (including next-gen Vera Rubin), AMD Instinct MI355X, Google Cloud TPU, and platforms such as Zhenwu 810
Architecturally, K3.1 combines DSA sparse attention (selecting up to 2048 key positions per query) with a recurrent Kimi Delta Attention (KDA) design; a single packed cache holds MLA latents, indexer keys, and KDA states together
Performance (vLLM relaying Kimi's internal benchmarks): K3.1 Pro scores 61.2% on Kimi Code Bench 3.0 Long-horizon, up more than 5 points from K3's 55.95%; the benchmark measures multi-step execution, tool use, and end-to-end engineering ability

Why it matters

Day-0 support means new models can be deployed on the leading open-source inference stack immediately at release, shrinking the gap between model launch and real-world usability
With train-inference consistency (especially for KDA) in place, RL training and online serving can share the same infrastructure—particularly valuable for long-horizon agent applications
If the 61.2% long-horizon coding score holds, it marks a generational gain for K3.1 on multi-step engineering tasks, though the figure comes from Kimi's internal benchmark and has yet to be independently verified

2026-11-05 ~ 2026-11-05 · 5 related posts

Primary sources

vLLM ships Day-0 support for Kimi K3.1: 61.2% on long-horizon coding bench, RL-ready stack — vllm_project · 2026-11-05 01:04 (UTC+08:00)
vLLM ships day-0 support for Kimi K3.1 across NVIDIA, AMD and TPUs — vllm_project · 2026-11-05 00:08 (UTC+08:00)
Kimi K3.1 mixes DSA with recurrent Kimi Delta Attention in one packed cache — vllm_project · 2026-11-05 00:07 (UTC+08:00)

[source] Kimi K3.1 mixes DSA with recurrent Kimi Delta Attention in one packed cache — vllm_project · 2026-11-05
Kimi K3.1 Pro hits 61.2% on long-horizon code bench, up from 55.95% — vllm_project · 2026-11-05
vLLM + Megatron add Day-0 RL support and long-agent serving for Kimi K3.1 — vllm_project · 2026-11-05
[source] vLLM ships day-0 support for Kimi K3.1 across NVIDIA, AMD and TPUs — vllm_project · 2026-11-05
[source] vLLM ships Day-0 support for Kimi K3.1: 61.2% on long-horizon coding bench, RL-ready stack — vllm_project · 2026-11-05