Kimi K3.1 Pro: What A Three-Point Frontier Gap Tells Us

🔍 Read the full analysis: Kimi K3.1 Pro: What A Three-Point Frontier Gap Tells Us on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Moonshot AI launched Kimi K3.1 Pro on Oct. 23, and independent evaluations from Artificial Analysis place the model fourth on its Intelligence Index at 55.1, three points behind the top score. The result puts the model near leading proprietary systems, but it is not open-weight yet: Moonshot says it plans to publish the weights by Nov. 4, and has not published a licence for K3.1.

Artificial Analysis’s Oct. 24 model-page figures put K3.1 Pro behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The evaluation compares models at their highest reasoning settings. Artificial Analysis describes K3.1 Pro as set to become the leading open-weights model once its weights are released; that ranking is conditional, not a description of a model users can download today.

The evaluation shows a particularly sharp result on Terminal-Bench 4.0, where K3.1 Pro scored 67%, compared with 13% for Kimi K3. Artificial Analysis also measured 1,846 Elo on GDPval-AA, a simulated knowledge-work evaluation, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, K3.1 Pro scored 52%, while Opus 5.5 scored 61%.

Artificial Analysis lists the model as a mixture of experts with 2.8 trillion total parameters and 104 billion active parameters, a one-million-token context window, and text and image input. Moonshot made it available through Kimi apps, Kimi Work, Kimi Code and its API. Artificial Analysis lists API pricing at $2 per million input tokens and $10 per million output tokens, estimating $2.11 per Intelligence Index task, versus $5.98 for Opus 5.5 at maximum reasoning.

A Narrower Gap, With Conditions

A score three points below the top model matters because open-weight systems can, once their weights and licence are available, give organisations more control over deployment than a hosted proprietary service. That can matter for data handling, customisation and the ability to run a model in a chosen environment. K3.1 Pro’s benchmark result suggests that this choice may come with a smaller performance trade-off than earlier comparisons showed.

But the benchmark result does not itself make the model open. Until the weights are published and the licence is known, customers cannot judge whether they may use, modify or redistribute them on the terms they need. The model’s scale also raises practical limits: running a system listed at 2.8 trillion total parameters is a substantial infrastructure undertaking, even though only 104 billion parameters are active at a time in the mixture-of-experts design.

Price is another point of comparison, not a guarantee of lower total cost. Artificial Analysis’s estimated cost per task is based on its evaluation setup; actual spending depends on prompts, output length, reasoning settings and usage patterns. The score and listed API rates offer a useful comparison, but they do not establish what a company’s full deployment will cost.

From Kimi K3 to K3.1 Pro

Kimi K3.1 Pro is described in the source material as a more heavily post-trained version of Kimi K3. In Artificial Analysis’s listed results, Kimi K3 scored 43.6 on the Intelligence Index, compared with 55.1 for K3.1 Pro. The same evaluation shows gains on agentic tasks: Terminal-Bench 4.0 rose from 13% to 67%. These are benchmark comparisons, not proof that every user task will improve by the same amount.

Moonshot’s launch materials also promote work involving game engines, parametric 3D modelling and finance, but those examples and several associated figures are vendor-reported. For instance, Moonshot says K3.1 Pro scored 92.7% on BenchCAD with tools, below the cited results for Opus 5.5 and GPT-6 Astra. It also reports a 65% reduction in claim-level hallucinations on its own Kimi Hallu Bench Hard test. Moonshot cautions that the test was deliberately designed to provoke errors and does not represent overall error rates in everyday use.

“The Kimi Hallu Bench Hard test is “deliberately crafted” to provoke errors.”

— Moonshot AI

Weights, Licence and Real-World Results

The main outstanding question is whether Moonshot will release the weights by its stated Nov. 4 target. The source material reports that no K3.1 licence had been published as of Oct. 24, so the permissions and restrictions governing use remain unknown. Until both details are available, the model’s open-weight status and practical accessibility cannot be assessed fully.

There is also a gap between independent benchmark scores and Moonshot’s demonstrations. Artificial Analysis independently measured several benchmark results, while claims about complete game projects, 3D workflows and finance performance rely on Moonshot’s own materials or internal tests. The source does not provide enough information to establish how those results generalise across independent users, tasks or deployments. Artificial Analysis’s reported hallucination rate fell from 53% to 44% on AA-Omniscience; that is distinct from Moonshot’s internal 65% reduction figure.

The Nov. 4 Release Test

The next milestone is Moonshot’s promised Nov. 4 publication of the weights. If it arrives, readers and developers will be able to check the release itself and examine the licence terms; those will determine whether the model can be run or adapted in their intended settings. The release would also allow independent researchers to test whether the benchmark performance holds up outside the hosted API.

Until then, K3.1 Pro can be assessed through its listed API and the available independent evaluation results, while Moonshot’s showcase claims should remain attributed to the company. A published licence, reproducible evaluations and evidence from use beyond launch demonstrations will clarify how much the close benchmark score changes the options available to organisations.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a model launched by Moonshot AI on Oct. 23, 2026. Artificial Analysis lists it as a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters.

How close is it to the top-ranked model?

Artificial Analysis scored K3.1 Pro at 55.1 on its Intelligence Index, three points below Claude Opus 5.5 at 57.6. The listed results put K3.1 Pro fourth overall.

Can users download the model weights now?

Not according to the information available as of Oct. 24. Moonshot has promised the weights by Nov. 4, but they had not yet been published, and no K3.1 licence was available in the source material.

Are the hallucination claims independently verified?

Moonshot’s claim of a 65% reduction comes from its internal Kimi Hallu Bench Hard test. Artificial Analysis reported a smaller change on its separate AA-Omniscience evaluation, from 53% to 44%; the tests measure different things and should not be treated as interchangeable.

What does the benchmark result mean for businesses?

It indicates that K3.1 Pro performed near leading proprietary models on Artificial Analysis’s evaluations. Businesses still need to consider the pending weights and licence, infrastructure requirements, actual API usage costs and whether the model performs well on their own tasks.

Source: ThorstenMeyerAI.com