Skip to content
audiotale.net
Sign in
NewsNews

Kimi K3.1 Pro launches: 55 on the Intelligence Index, open weights due by 4 November

Moonshot's updated open-weight model ranks fourth on Artificial Analysis's index and is on sale now; the weights, and their licence, are still to come.

audiotale.net team

· 6 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What to do
  3. What happened
  4. Background
  5. The numbers
  6. How the cyber data was handled
  7. What it means
  8. What we don't know yet
  9. What to watch next
  10. Sources

Moonshot AI launched Kimi K3.1 Pro on 23 October 2026, a post-trained update to its open-weight Kimi K3 that, according to independent tester Artificial Analysis, scores 55 on its Intelligence Index and ranks fourth overall, behind only Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full. The catch for anyone who wants to run it themselves: the weights are not published yet. Moonshot says they will follow by 4 November 2026.

Key points

  • Kimi K3.1 Pro is the first model in Moonshot's K3.1 series, a 3T-class open-weight model with a 1-million-token context window; it is available now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API.
  • Artificial Analysis gives it 55 on its Intelligence Index, 11 points above Kimi K3 (44), placing it fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56).
  • Moonshot says the full weights will be released by 4 November 2026; Artificial Analysis says it will lead open-weights models once they are published.
  • API prices are $2.00 per million input tokens and $10.00 per million output tokens, about a third below K3, according to Artificial Analysis.
  • Moonshot says it removed cyber-related data from mid-training and reinforcement-learning stages for the open release; its own test shows exploit success falling from 90.5% to 81.0%.

What to do

  • Use it today: pick kimi-k3.1-pro or kimi-k3.1-pro-highspeed in the Kimi API Platform, or run /model in Kimi Code and select Kimi K3.1 Pro. Kimi Work needs desktop app version 3.1.0 or later (Windows and Apple-silicon Macs).
  • Run it yourself: you cannot yet. Wait for the weights. We have not seen a licence for K3.1 Pro; the earlier K3 repository on Hugging Face uses a bespoke "Kimi K3 License", so read the licence that ships with the new weights before building on it.
  • Enterprise admins: Moonshot says access is off by default at launch and must be enabled by an administrator.
  • Check hardware first: K3 is a 2.8-trillion-parameter model with 104 billion active parameters per token, per its Hugging Face page, so self-hosting means a multi-GPU server, not a workstation.

What happened

Moonshot announced Kimi K3.1 Pro on its blog as "our most capable 3T-class open-weight model to date". It is a post-trained version of Kimi K3 with, the company says, substantially larger post-training compute, built around large-scale joint reinforcement learning across hundreds of downstream domains. A second variant, K3.1 Pro Highspeed, is meant to be 2-3 times faster.

The company's headline coding result is Terminal-Bench 4.0, where it reports 65.7% for K3.1 Pro, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%). Artificial Analysis, which ran its own evaluations, lists 67% for K3.1 Pro on the same benchmark (up from 13% for K3) and says that is the highest of any model it tested; it lists Claude Sonnet 5.5 at 64%. The two figures differ because the harnesses and settings differ, so treat them as two measurements, not one.

Other numbers from Moonshot's own benchmarks: 92.7% on BenchCAD with tools (Claude Opus 5.5: 96.2%, GPT-6 Astra: 95.9%) and 64.4 on OfficeQA Pro v2. On its internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot stresses that this is a stress test deliberately built to provoke errors and does not represent everyday error rates. Most of these are company-reported results.

Background

Kimi K3 was Moonshot's previous flagship: a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token, a vision encoder, and a context window of 1,048,576 tokens, according to its Hugging Face page. It uses MXFP4 weights with MXFP8 activations through quantization-aware training. The page recommends vLLM, SGLang and TokenSpeed for serving, and its code and weights fall under the Kimi K3 License.

K3.1 Pro follows the same pattern Moonshot used for K3: launch through its own products and API first, publish weights afterwards. Artificial Analysis and The Decoder both describe K3.1 Pro as having the same 2.8T total and 104B active parameter count, which fits a post-trained update of K3 rather than a new base model; Moonshot's blog does not list the figures itself.

The numbers

Artificial Analysis's composite Intelligence Index puts the new model within three points of the leader. If the weights ship as promised, the gap between the best open-weights model and the best proprietary one shrinks from 12 points to 3, according to the same source.

Artificial Analysis Intelligence Index, top models and Kimi K3index points
Artificial Analysis Intelligence Index, top models and Kimi K3 (index points)
index points
Claude Opus 5.558
Claude Sonnet 5.556
Muse Wonder Full56
Kimi K3.1 Pro55
Kimi K344

Source: Artificial Analysis, 23 October 2026

On cost, Artificial Analysis reports about $1.98 per Intelligence Index task, roughly the same as K3, even though per-token prices fell by about a third: K3.1 Pro produced about 14% more output tokens per task (55K versus 48K). Its listed API price is $2 input, $10 output and $0.20 for cached input per million tokens. The Decoder reports Highspeed at $4 input and $20 output.

Artificial Analysis also reports 1,846 Elo on GDPval-AA, third place and 312 above K3, against 1,866 for Claude Opus 5.5 and 1,857 for Muse Wonder Full. Its Omniscience hallucination rate for K3.1 Pro is 44%, down from 53%.

How the cyber data was handled

The unusual part of the announcement is a long section on cybersecurity. Moonshot says vulnerability-research ability kept rising as post-training scaled, and that it traced this to cyber-related content in the general mid-training corpus and cyber tasks in the reinforcement-learning mix. For the open-weight release it removed that data from both stages.

Its internal benchmark, xdev-bench-hard, asks an agent to turn real, recently disclosed bugs (in software such as the Linux kernel, Node.js and JavaScriptCore) into working exploits. A checkpoint without the removals solves 90.5% of problems at pass@3; the final K3.1 Pro solves 81.0%. The Decoder reports that DeepSeek-V4.1-Flash scores 85.7% on the same test and K3 scored 33.3%. These are Moonshot's own results on its own benchmark and have not been independently reproduced.

Moonshot's stated policy has two parts: not to release open weights that would extend the absolute frontier of cyber capability, and to raise the capability available in open weights gradually, arguing that open models let more teams audit code, including legacy systems.

What it means

For developers, a model near the top of an independent index with open weights on the way is the main news: it could be run on private infrastructure, fine-tuned and audited, subject to the licence. The price cut makes the hosted API a cheaper way to test it before then, though higher token use erodes some of the saving.

For the proprietary labs, the gap is now three index points, not twelve, on this measure. For security teams, the cyber-data removal is a concrete, quantified attempt to limit what an open model can do, and it will be tested by outside researchers once the weights are available.

What we don't know yet

  • The licence. Neither Moonshot's blog nor Artificial Analysis states the terms for K3.1 Pro's weights.
  • Whether weights arrive on time. Moonshot says "by Nov 4, 2026"; Artificial Analysis gives no date.
  • Independent replication of the company-reported figures, including Hallu Bench, BenchCAD and the cyber test.
  • Hardware requirements for the new weights; we have only K3's published architecture to go on.

What to watch next

Look for the Hugging Face repository under the moonshotai organisation on or before 4 November, its licence file, serving recipes from vLLM and SGLang, and independent evaluations, including whether the cyber-data removal holds up under outside testing.

Sources

SaveStart the discussion

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.

Read next