Kimi K3.1 Pro: An Open-Weight Giant Comes Within Three Points Of The Frontier

🔍 Read the full analysis: Kimi K3.1 Pro: An Open-Weight Giant Comes Within Three Points Of The Frontier on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get movie nights 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 in its apps, coding product and API on Oct. 23, and an evaluation by Artificial Analysis gave it a score of 55.1, placing it fourth on the group’s Intelligence Index. The result puts the model three points behind the index leader, but its weights are not yet available: Moonshot says it plans to publish them by Nov. 4, and has not published a licence for the release.

Artificial Analysis lists K3.1 Pro as a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters. It supports text and image input and has a 1-million-token context window. The company’s services went live Oct. 23 through Kimi apps, Kimi Work, Kimi Code and the Kimi API.

On Artificial Analysis’s Intelligence Index, K3.1 Pro scored 55.1. Claude Opus 5.5 scored 57.6, while Claude Sonnet 5.5 and Muse Wonder Full scored 56.0 and 55.8. Artificial Analysis said K3.1 Pro would lead open-weight models once its weights are published; that ranking is conditional, because the weights have not been released.

The listed API price is $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimated $2.11 per Intelligence Index task for K3.1 Pro, compared with $5.98 for Opus 5.5 at the evaluated settings. Those figures describe the evaluation’s task costs, not a guarantee of the price for every use case.

A Narrower Gap, With a Caveat

The score suggests that a model whose weights users can inspect or run locally could approach leading proprietary systems on this particular aggregate benchmark. If the promised release occurs under a usable licence, developers and companies may gain more control over deployment, data handling and modification than they get from an API-only product. But open-weight availability is not yet a present feature: users cannot assess the actual weights or run them independently until they are published.

The price comparison may also draw interest from organizations weighing model performance against inference costs. Artificial Analysis’s task estimate puts K3.1 Pro below Opus 5.5 on cost, while the benchmark scores are relatively close. Real costs will depend on workload, serving infrastructure and the amount of reasoning or output required. The model’s large listed parameter count also means self-hosting could demand substantial data-centre resources, rather than ordinary workstation hardware.

Artificial Analysis reported a sharp rise on its Terminal-Bench 4.0 agentic coding test: 67% for K3.1 Pro, against 13% for K3. On GDPval-AA, a simulated knowledge-work evaluation, it scored 1,846 Elo, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. These are benchmark results, not proof that the model will perform equally well on every company’s tasks.

From K3 to K3.1 Pro

K3.1 Pro is described as a more heavily post-trained version of Kimi K3. Artificial Analysis’s reported score for K3 was 43.6, compared with 55.1 for the new model. The same evaluation showed a smaller increase on Humanity’s Last Exam, from 47% to 52%; Opus 5.5 scored 61%. On Artificial Analysis’s AA-Omniscience hallucination evaluation, K3.1 Pro recorded 44%, compared with 53% for K3. The results indicate gains across some measured areas, but the size of the improvement varies by test.

Moonshot has also presented demonstrations and internal test results for work involving game engines, finance and 3D modelling. It says K3.1 Pro can create projects in Unity, Unreal Engine and Godot, and generate parametric models for Blender and CAD tools. The company reports a 92.7% result on BenchCAD with tools, below the figures it cites for Claude Opus 5.5 and GPT-6 Astra. These examples and internal evaluations are Moonshot’s claims; they should not be confused with the separate Artificial Analysis results.

Moonshot’s claimed 65% reduction in hallucinations came from its internal Kimi Hallu Bench Hard test, where its claim-level rate fell from 23.4% for K3 to 8.3% for K3.1 Pro. The company says the test was deliberately designed to provoke errors and that its rates do not represent overall error levels in daily use. Artificial Analysis measured a smaller change on its own hallucination evaluation, from 53% to 44%.

“The internal hallucination test is “deliberately crafted” to provoke errors, and its rates “do not represent overall error rates in daily use.””

— Moonshot AI

Weights, Licence and Real-World Results

The main outstanding question is whether Moonshot will publish the weights by its stated Nov. 4 target. The source material does not establish that the files are available now. It also reports no published K3.1 Pro licence, leaving unclear what users would be allowed to do with the weights, including commercial use or redistribution.

Independent benchmark scores offer a basis for comparison, but they do not settle how the model performs across everyday workflows. Moonshot’s demonstrations and internal benchmarks have not been independently verified in the supplied material, and the company’s own hallucination-test caveat warns against treating the reported rates as general error rates. The hardware and operating costs of running the full model locally are also not specified.

The Nov. 4 Release Test

The next milestone is Moonshot’s stated plan to publish K3.1 Pro’s weights by Nov. 4. A release would allow outside users to inspect the files, test them on their own systems and review the licence terms. Those details will determine whether the model can be used as an open-weight alternative in practice, and under what conditions.

Further comparisons will depend on independent testing after the weights become available, along with more evidence from real deployments. Until then, the confirmed picture is a commercially available model with a strong Artificial Analysis score, alongside an open-weight release that remains promised rather than completed.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a Moonshot AI model released in the company’s apps, Kimi Work, Kimi Code and API on Oct. 23, 2026. Artificial Analysis lists it as a 2.8-trillion-total-parameter mixture-of-experts model with 104 billion active parameters.

How did it score against leading models?

Artificial Analysis scored it 55.1 on its Intelligence Index, placing it fourth in the reported comparison. Claude Opus 5.5 led at 57.6. The index is one evaluation framework and does not establish performance on every task.

Can users download the weights now?

The source material says Moonshot promised the weights by Nov. 4; it does not report that they have been released. No K3.1 Pro licence had been published in the material available for this report.

Is the 65% hallucination reduction independently confirmed?

No. That figure comes from Moonshot’s internal Kimi Hallu Bench Hard test. Artificial Analysis reported a smaller change on its own evaluation, from 53% for K3 to 44% for K3.1 Pro.

How much does K3.1 Pro API use cost?

The listed price is $2 per million input tokens and $10 per million output tokens. Actual costs depend on usage, and the source’s estimate of $2.11 per Intelligence Index task applies to its benchmark methodology.

Source: ThorstenMeyerAI.com