Kimi K3.1 Pro Is Three Points Off The Frontier, Without Released Weights

🔍 Read the full analysis: Kimi K3.1 Pro Is Three Points Off The Frontier, Without Released Weights on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Moonshot AI launched Kimi K3.1 Pro on October 23, with the model reaching 55.1 on Artificial Analysis’s Intelligence Index—3 points below the top score in the comparison. The launch puts a model Moonshot describes as open-weight close to leading proprietary systems, but its weights have not been released and the company has not published a K3.1 licence.

Artificial Analysis’s October 24 model-page figures place Kimi K3.1 Pro fourth overall, behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0, and Muse Wonder Full at 55.8. The rankings reflect the models at their highest reasoning settings, where reported. Artificial Analysis describes K3.1 Pro as a prospective leading open-weights model, conditional on the weights being published; it is not currently downloadable on that basis.

Artificial Analysis lists K3.1 Pro as a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters, a one-million-token context window, and text and image input. Moonshot says it is a more heavily post-trained version of Kimi K3. The model became available through the Kimi apps, Kimi Work, Kimi Code and the Kimi API. Listed API rates are $2 per million input tokens and $10 per million output tokens.

On Artificial Analysis’s measure of cost per Intelligence Index task, K3.1 Pro is listed at $2.11, compared with $5.98 for Claude Opus 5.5 at maximum effort. The cost figure is tied to that evaluation measure; it is not a general estimate of every user’s bill. Artificial Analysis also reports a Terminal-Bench 4.0 score of 67% for K3.1 Pro, against 13% for Kimi K3 in its comparison. These are independent evaluation results, distinct from Moonshot’s own product demonstrations and internal tests.

A Smaller Gap, With a Condition

The score makes Kimi K3.1 Pro relevant to organisations weighing model capability, cost and control. Its result is close to the highest-scoring systems in this comparison, while its listed API price is lower than Opus 5.5’s. The Terminal-Bench result also suggests a substantial improvement over the previous Kimi K3 entry on that particular test. Those results may prompt developers to evaluate it for coding and agentic tasks rather than treating it as a general-purpose alternative based only on vendor claims.

The possibility of running the model independently matters for companies with data-control, deployment or customization requirements. But that possibility is not yet a practical option: the weights are pending, and the licence is unpublished. Artificial Analysis’s ranking measures benchmark performance, not ease or cost of self-hosting. At 2.8 trillion total parameters, deployment would require substantial infrastructure; the source describes self-hosting as a data-centre-scale project. The score alone does not establish that an organisation can run it economically or use it under terms it finds suitable.

From Kimi K3 to K3.1

Moonshot presents K3.1 Pro as a post-trained update to Kimi K3, rather than a wholly separate model family. In Artificial Analysis’s comparison, the earlier K3 scored 43.6 on the Intelligence Index, while K3.1 Pro scored 55.1. The same evaluation lists their Terminal-Bench 4.0 results at 13% and 67%, respectively. The size of that difference is specific to the named benchmark and should not be treated as proof of a similar improvement across every task.

Moonshot’s launch materials describe examples involving game engines, Blender and CAD, finance work, and reduced hallucinations. Those are company-reported demonstrations and tests, not all independently verified findings. For example, Moonshot reports a 92.7% BenchCAD score with tools, compared with 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. It also reports a 65% reduction in claim-level hallucinations on its internal Kimi Hallu Bench Hard test. The source says Artificial Analysis measured a smaller change on its own hallucination evaluation, from 53% to 44%, underscoring that results depend on the test used.

“Kimi K3.1 Pro is its most capable “3T-class open-weight model.””

— Moonshot AI

Weights, Licence and Real-World Results

The most immediate open question is whether Moonshot will publish the weights by its stated November 4 target. Until then, outside users cannot verify the model by downloading and running those weights. The company has also not published a K3.1-specific licence, leaving the permitted uses and conditions unclear. The source material does not establish whether the announced schedule or licence terms may change.

Benchmark scores do not settle how the model will perform across production workloads, languages, tools or long-running tasks. Moonshot’s game-development, CAD and finance examples are vendor demonstrations, while its hallucination claim comes from an internal stress test that the company says is not representative of ordinary use. Artificial Analysis’s result—an AA-Omniscience rate falling from 53% to 44%—is a separate measure, not confirmation of Moonshot’s 65% figure. The evidence therefore supports a strong benchmark showing, but not a blanket conclusion about reliability or everyday performance.

It is also unclear what hardware, operating costs and engineering work would be required for a particular organisation to self-host the full model. API prices offer a point of comparison for use through Moonshot’s service, but they do not answer the cost question for local deployment.

November 4 Release Target

The next stated milestone is Moonshot’s promised publication of Kimi K3.1 Pro’s weights by November 4. A release would allow developers and independent evaluators to test the model outside the company’s hosted products, subject to the terms Moonshot publishes. The licence, model files and practical deployment requirements will determine whether it can be used as an open-weight option in more than name.

Readers and potential users can also watch for further independent evaluations, especially results on the tasks they care about and tests of the downloadable model if it appears. Until those details are available, the clearest confirmed picture is the launch, the Artificial Analysis scores and the current API rates; self-hosting access and K3.1 usage terms remain unresolved.

Key Questions

What is Kimi K3.1 Pro?

It is a model launched by Moonshot AI on October 23, 2026. Artificial Analysis lists it as a mixture-of-experts system with 2.8 trillion total parameters, 104 billion active parameters, a one-million-token context window, and text and image input.

How close did it come to the top score?

Artificial Analysis scored Kimi K3.1 Pro at 55.1 on its Intelligence Index. The top listed score was 57.6 for Claude Opus 5.5, a difference of 2.5 points—described in the source headline as about three points.

Can users download and run the weights now?

No. Moonshot says the weights are due by November 4. As of the October 24 source report, they had not been released, and no K3.1 licence had been published.

What does Kimi K3.1 Pro cost through the API?

The listed rates are $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimates $2.11 per Intelligence Index task, a benchmark-specific cost measure rather than a prediction of every user’s bill.

Are Moonshot’s coding and hallucination claims independently confirmed?

Not all of them. Moonshot’s demonstrations and internal tests are company-reported. Artificial Analysis independently evaluated the model and reported its own benchmark and hallucination results, which do not match every Moonshot figure or establish performance in all real-world settings.

Source: ThorstenMeyerAI.com