What happened

On 23 October 2026, Moonshot AI, the Chinese lab behind the Kimi models, released Kimi K3.1 Pro. It is available now in the Kimi apps and through Moonshot’s API under the model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot describes it as its most capable open-weight model, but the weights are not public yet: according to Simon Willison, quoting Moonshot’s announcement, “full model weights will be released by Nov 4, 2026.”

The model is a post-trained version of Kimi K3 with a 1-million-token context window. Moonshot’s launch post is the primary source for this story. We read the launch post and the API quickstart on 25 October 2026; the pages do not state exact parameter counts or dollar prices, and the benchmark tables did not come through in our reading. Figures that are not on those pages (size, price, the Artificial Analysis score and some benchmark numbers) come from MarkTechPost, The Next Web and Simon Willison and are attributed to them. We will update this article once the Hugging Face repository and Moonshot’s pricing table can be checked directly.

Key details

  • Size (not on Moonshot’s pages; as reported by MarkTechPost): a sparse mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token and multimodal input. Moonshot itself only calls it “3T-class” and confirms the 1-million-token context window.
  • Price on Moonshot’s API (the pricing table was empty in our reading; MarkTechPost and Simon Willison report these figures): $2.00 per million input tokens, $0.20 for cache hits and $10.00 per million output tokens. A Highspeed variant that Moonshot says is two to three times faster costs $4.00 and $20.00. The earlier Kimi K3 was priced at $3.00 and $15.00.
  • One independent score: Artificial Analysis rates Kimi K3.1 Pro at 55 on its Intelligence Index, behind Claude Opus 5.5 (58) and ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (53 each), according to Simon Willison.
  • Moonshot’s own benchmarks (not independently verified): on Terminal-Bench 4.0 Moonshot’s page reports 65.7% for K3.1 Pro against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Press coverage also cites 51.7 on SWE-Marathon against 46.3 for Opus 5.5, and 38.1 on Terminal-Bench Science 0.1 against 68.1 for GPT-6 Astra; we could not read those numbers on the page. The comparison figures for rival models come from Moonshot’s chart and were not confirmed on those companies’ pages.
  • Hallucination and cyber claims (Moonshot’s internal tests, on its page): Moonshot says the claim-level hallucination rate on its Kimi Hallu Bench Hard fell from 23.4% for K3 to 8.3% with tools. It also says it “identified and removed cyber-related data from both training stages”: on its in-house xdev-bench-hard (pass@3) the shipped model scores 81.0%, against 90.5% for a checkpoint without that data removal and 33.3% for K3.
  • Not yet known: the license for the weights, the platform they will be published on, and the memory needed to run them. MarkTechPost notes that K3 needed supernodes of 64 or more accelerators; this is not a model for a single GPU.

Why it matters

If the weights ship on schedule, K3.1 Pro would be one of the largest openly downloadable models, and a lab outside the US would again be the one setting the pace for open weights. Artificial Analysis’s score of 55, within a few points of the best closed model, is the most reliable signal in the launch, because it does not depend on Moonshot’s own tests.

For people who use AI tools rather than host models, the practical news is the price: $2 per million input tokens is a third lower than K3, and Simon Willison notes it is half of Claude Opus 5.5’s price. Reading the numbers needs care. The coding and agent results are self-reported and mixed, the hallucination figure comes from an internal test, and “open” will mean little until the license is published. The cyber-capability reduction is an unusual step for an open-weight release and will be worth checking when outside evaluators can test the weights.

Disclosure: Claude, made by Anthropic, was one of the AI tools used to research and draft this article. Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 are compared with Kimi K3.1 Pro in the figures above. Performance claims are Moonshot’s own unless stated otherwise. Weights timing, context length, model IDs, hallucination and cyber figures and Terminal-Bench 4.0 were checked against Moonshot’s pages; size, price and the Artificial Analysis score were not.

Sources

  1. Introducing Kimi K3.1 Pro (Moonshot AI blog) Primary source , Moonshot AI
  2. Kimi K3.1 Pro API quickstart Primary source , Moonshot AI
  3. Introducing Kimi K3.1 Pro Simon Willison
  4. Moonshot AI releases Kimi K3.1 Pro: a 2.8T total / 104B active params (MoE) model with 1M context MarkTechPost
  5. Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills The Next Web