🔍 Read the full analysis: Kimi K3.1 Pro: An Open-Weight Giant Lands Three Points From The Frontier. The Weights Aren’t Out Yet on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Moonshot AI released Kimi K3.1 Pro in its apps and API on October 23. Artificial Analysis scored it 55.1, placing it fourth in its Intelligence Index, while Moonshot says the model’s weights will be published by November 4; the licence has not been announced.
Moonshot AI launched Kimi K3.1 Pro on October 23, with access through the Kimi apps, Kimi Work, Kimi Code and its API. Artificial Analysis scored the model 55.1 on its Intelligence Index, ranking it fourth behind three proprietary models; Moonshot says it will publish the model weights by November 4, but has not yet published a licence for them.
Artificial Analysis lists K3.1 Pro as a mixture-of-experts model with 2.8 trillion total parameters, of which 104 billion are active, and a one-million-token context window. It accepts text and images. The Intelligence Index result places it three points below the listed leader, Claude Opus 5.5 at 57.6, and just below Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. These scores reflect the models and settings in Artificial Analysis’s October 24 comparison.
The same evaluation puts K3.1 Pro’s estimated cost at $2.11 per Intelligence Index task, compared with $5.98 for Opus 5.5 at its highest reasoning setting. Artificial Analysis lists K3.1 Pro’s API prices as $2 per million input tokens and $10 per million output tokens. These are evaluation-based cost estimates, not a guarantee that every customer’s workload will cost the same.
On specific tests, Artificial Analysis measured a large gain in agentic performance: 67% on Terminal-Bench 4.0, compared with 13% for the earlier Kimi K3. It reported a 1,846 Elo score on GDPval-AA, a simulated knowledge-work evaluation, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. Results vary by benchmark: K3.1 Pro scored 52% on Humanity’s Last Exam, while Opus 5.5 scored 61%.
Infographic 01 · Claimed vs. measured
Moonshot says “matches the strongest.” Independent tests say: close.
The KPIs at the top are independent measurements. Vendor numbers and independent numbers come from different harnesses, so they are shown side by side, never mixed.
Moonshot reports
vendor claims own harness, partly internal tests
| Benchmark | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | vs. Claude Opus 5.5: 66.4% |
| BenchCAD with tools | 92.7% | vs. Claude Opus 5.5: 96.2% |
| OfficeQA Pro v2 | 64.4 | vs. Claude Opus 5.5: 62.2 |
| Kimi Hallu Bench Hard internal | 8.3% | vs. Kimi K3: 23.4% (about −65%) |
Artificial Analysis measured
independent highest reasoning setting per model
| Measurement | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 67% | vs. Claude Opus 5.5 (max): 60% |
| GDPval-AA | 1,846 Elo | vs. Claude Opus 5.5: 1,866 |
| Humanity’s Last Exam | 52% | vs. Claude Opus 5.5: 61% |
| AA-Omniscience hallucination rate | 44% | vs. Kimi K3: 53% · Gemini 4 Argon: 15% |
The 65% hallucination drop holds on Moonshot’s own stress test. On an independent knowledge test, the rate fell from 53% to 44%.Moonshot: the stress test rates “do not represent overall error rates in daily use.”
Sources: Moonshot AI, Kimi K3.1 Pro tech blog (23 Oct 2026) (left card, vendor charts); Artificial Analysis launch article (23 Oct 2026) and Artificial Analysis model page (data as of 24 Oct 2026) (KPIs and right card).
Powered by Thorsten Meyer AI
A Narrower Gap for Open Weights
If Moonshot releases usable weights under terms that permit the intended use, K3.1 Pro could give developers and organisations an option to run a high-scoring model on infrastructure they control. The benchmark result is notable because Artificial Analysis places it three points behind its top-ranked model, narrowing the measured gap between this model and the proprietary systems in the comparison.
That possibility is not the same as present-day open access. The weights are not yet available, the licence remains unpublished, and running a model listed at 2.8 trillion total parameters would require substantial infrastructure. Until those conditions are clearer, most users can access K3.1 Pro through Moonshot’s hosted products and API rather than download and operate it themselves.
Price may matter to businesses comparing hosted models, but benchmark cost estimates cannot settle the full purchasing decision. Performance on a company’s own tasks, reliability, privacy requirements, service availability and operating costs all affect whether a model is a practical alternative. The published scores offer a comparison point, not proof that K3.1 Pro will perform as well in every use case.
GPU server for running large language models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Kimi K3 to K3.1 Pro
K3.1 Pro is described by Moonshot as a more heavily post-trained version of Kimi K3. In the Artificial Analysis comparison supplied for October 24, the earlier Kimi K3 scored 43.6 on the Intelligence Index, while K3.1 Pro scored 55.1. The new model’s score also exceeds the listed results for open-weight models available at the time, including Xiaomi’s MiMo-V2.6-Pro at 46.3 and DeepSeek V4.1 Flash at 45.1.
Artificial Analysis describes K3.1 Pro as set to lead the open-weight category once its weights are published. That qualification matters: the score has been measured, but the model is not yet downloadable as open weights. Whether the eventual release matches the hosted version, and whether its licence permits commercial use or redistribution, cannot be established from the information available at launch.
Moonshot’s launch material also presents examples involving game engines, 3D modelling and finance. Those are company demonstrations and reported internal results, not independent confirmation of broad capability. The independent benchmark results provide a separate basis for comparison, although they too represent performance on specified tests rather than every real-world task.
“The Kimi Hallu Bench Hard test is deliberately crafted to provoke errors, and the rates do not represent overall error rates in daily use.”
— Moonshot AI
Weights, Licence and Real-World Results
The central open question is whether Moonshot will publish the weights by its stated November 4 target, and what licence will govern them. No K3.1 Pro licence had been published in the source material. Without those details, users cannot determine the permitted uses, redistribution terms or restrictions. The promised release also remains a future commitment, not a completed event.
Some prominent claims come from Moonshot’s own tests. The company says its internal Kimi Hallu Bench Hard claim-level hallucination rate fell from 23.4% for Kimi K3 to 8.3% for K3.1 Pro, describing that as a 65% reduction. Artificial Analysis measured a smaller change on its own AA-Omniscience evaluation, from 53% to 44%. The tests use different methods, so the figures should not be treated as directly interchangeable or as a universal rate of errors in everyday use.
Moonshot also reports results for game development, parametric modelling and finance, including a 92.7% result on BenchCAD with tools. These claims are based on the company’s reported tests and demonstrations. The available material does not establish how often the model can reproduce those results in independent evaluations or production settings.
The November Weights Deadline
The next milestone is Moonshot’s promised publication of K3.1 Pro’s weights by November 4. A release would allow independent observers to check whether the downloadable model is the same version tested by Artificial Analysis and to assess the practical hardware demands. The licence will be just as consequential for developers deciding whether they can adapt, deploy or redistribute the model.
Until then, users can test the hosted model through Moonshot’s products and API, subject to the company’s service terms and pricing. Further independent evaluations could clarify how its benchmark gains translate to coding, agentic workflows and knowledge work outside the published test suites. The timing and scope of any such evaluations are not yet confirmed.
Key Questions
When did Kimi K3.1 Pro launch?
Moonshot AI made Kimi K3.1 Pro available on October 23, 2026, through the Kimi apps, Kimi Work, Kimi Code and its API.
How did Kimi K3.1 Pro rank?
Artificial Analysis gave it a 55.1 Intelligence Index score, ranking it fourth in the comparison as of October 24. Claude Opus 5.5 scored 57.6, Claude Sonnet 5.5 scored 56.0 and Muse Wonder Full scored 55.8.
Can users download the model weights now?
No. Moonshot says it plans to publish the weights by November 4. They had not been released in the source material, and no licence had been published.
Are the hallucination and coding claims independently verified?
Some performance data comes from Artificial Analysis, including its benchmark scores and its AA-Omniscience result. Claims about internal hallucination tests and demonstrations across game engines, 3D modelling and finance come from Moonshot’s own material and should be treated as vendor-reported results.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
