FULL STORY

Kimi K3.1 Pro Launches Open-Source to a Lukewarm Reception

Moonshot released Kimi K3.1 with day-0 support from vLLM and SGLang, then open-sourced the 2.8T-parameter K3.1 Pro. Despite the launch, community buzz stayed unusually low, with users questioning whether past paid-promotion controversies played a role.

2026-11-05 ~ 2026-11-07 · 3 episodes · 18 posts

Episode 1 · Kimi K3.1 gets Day-0 support from vLLM and SGLang across NVIDIA, AMD and TPU (2026-11-05, 10 posts)

On launch day of Moonshot's Kimi K3.1, both vLLM and SGLang announced Day-0 support spanning serving, RL training, and multiple hardware platforms, with Modal offering online inference from day one. Kimi's internal benchmark shows K3.1 Pro scoring 61.2% on long-horizon coding, up more than 5 points over K3. With hardware vendors and third-party teams joining the effort, the 2.8T-parameter MoE model became deployable immediately upon release.

Confirmed

  • vLLM and inferact shipped a vLLM + Megatron combination providing Day-0 RL support for Kimi K3.1, covering train-inference consistency for Kimi Delta Attention and long-horizon agent serving.
  • Hardware coverage is broad: NVIDIA GPUs (including next-gen Vera Rubin), AMD Instinct MI355X, Google Cloud TPU, and the Zhenwu 810 platform.
  • Google partnered with Inferact (vLLM) and RadixArk (SGLang) to deliver Day-0 distributed inference for the 2.8T-parameter MoE Kimi K3.1 Pro on TPU v7x; vLLM announced its llm-d solution as available. This is the first time a Moonshot flagship model has received Day-0 TPU inference support.
  • The SGLang team, with RadixArk, NVIDIA, AMD, Moonshot, and SemiAnalysis, completed Day-0 inference support for Kimi K3.1 Pro, publishing deployment docs and Cookbook recipes; GB300 delivers 31K tok/s single-card throughput at high concurrency, with the megakernel available out of the box.
  • Modal announced a bring-up collaboration with SGLang—validating checkpoint accuracy and optimizing the serving path—so Kimi K3.1 Pro was servable on Modal from day zero.
  • Architecturally, K3.1 combines DSA sparse attention (each query selects at most 2048 key positions) with recurrent Kimi Delta Attention (KDA); a single packed cache holds MLA latents, indexer keys, and KDA states.
  • Performance (vLLM relaying Kimi's internal benchmark): K3.1 Pro scores 61.2% on Kimi Code Bench 3.0 Long-horizon versus 55.95% for K3, a gain of over 5 points; the benchmark tests multi-step execution, tool use, and end-to-end engineering.

Unconfirmed

  • The 61.2% long-horizon coding score comes from Kimi's internal benchmark and lacks independent third-party verification.

Why it matters

  • Day-0 support means a new model is deployable on mainstream open-source inference stacks immediately; this round's dual-track adaptation by vLLM and SGLang further shrinks the gap between model release 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.
  • K3.1 Pro's first-ever Day-0 TPU support shows Google actively pushing third-party frontier models onto its own silicon; NVIDIA and AMD's simultaneous participation reflects that frontier-model launch day has become a competitive battleground for inference stacks and hardware ecosystems.
  • If the 61.2% long-horizon coding figure holds, it signals a generational gain for K3.1 on multi-step engineering tasks.

Episode 2 · Moonshot's Kimi K3.1 Pro Launches Across Multiple Platforms (2026-11-05, 6 posts)

Moonshot AI has released Kimi K3.1 Pro: an open-weight MoE model with 2.8 trillion total parameters and 104B activated parameters. Weights are already up on Hugging Face. The model is built for agentic use cases, supports text and image understanding, and offers a context window of up to 1 million tokens — positioned as an upgrade to the K3 series for coding, research, and other multi-step tasks.

Confirmed

  • Per the ThursdAI podcast account, K3.1 Pro is a 2.8T-total-parameter MoE with 104B activated parameters; weights are on Hugging Face, though specific benchmark scores were not yet available at the time
  • The model is available on Microsoft Foundry with up to 1 million tokens of context (relayed by @DanWahlin)
  • You.com Inference Router offers Day 0 support (relayed by @PolarBearby)
  • It is live on the Tinker platform with 256K context; Tinker calls it the strongest open-weight model on multiple agentic coding and knowledge-work benchmarks, and its largest model to date (@kevinlu)
  • SGLang announced Day-0 support: the model replaces dense MLA with sparse attention (DSA), reaching up to 33.7K tok/s per card (AgentX, GB300)
  • @AnushElangovan (NVIDIA) posted "Day 0: K3.1 Pro" with inference TCO curves from SemiAnalysis InferenceX (updated 2026-11-06), noting 45 million tokens per dollar on MI355X

Why it matters

  • Open weights combined with a 1-million-token context and an agentic focus put it directly in contention for agentic coding and long-context workloads
  • The DSA sparse-attention design delivers higher serving efficiency and a lower deployment barrier; SGLang's Day-0 support and third-party TCO measurements show the ecosystem and hardware adaptation are moving fast
  • Rapid rollouts across Microsoft Foundry, You.com, Tinker, and other platforms in a short window reflect how much distribution channels value the model

Episode 3 · Kimi K3.1 Pro Meets Silence After Open-Source Release, Ranks 7th on autoresearch (2026-11-07, 2 posts)

Kimi K3.1 Pro has seen unusually low discussion after its open-source release, possibly due to Moonshot's past sponsorship controversy, and ranks 7th on Josh Purtell's autoresearch benchmark behind GPT-6 Astra at 72.6%.