Open-Weight Giant Kimi K3.1 Pro Lands Near The Frontier

🔍 Read the full analysis: Open-Weight Giant Kimi K3.1 Pro Lands Near The Frontier 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 October 23, and an evaluation by Artificial Analysis gave it an Intelligence Index score of 55.1, placing it fourth among the models listed and close to leading proprietary systems. The model is available through Moonshot’s apps and API, but its weights are not yet public: the company says it plans to release them by November 4, and has not published a licence for K3.1 Pro.

Artificial Analysis lists K3.1 Pro as a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters. It accepts text and images and has a 1-million-token context window. Moonshot describes it as a more heavily post-trained version of Kimi K3. The launch is available through Kimi apps, Kimi Work, Kimi Code and the Kimi API.

In Artificial Analysis’s figures as of October 24, K3.1 Pro scored 55.1 on the Intelligence Index, behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The same evaluation puts its cost at about $2.11 per index task, compared with $5.98 for Opus 5.5 at its highest reasoning setting. Those prices and rankings reflect the benchmark’s listed configurations, not a guarantee of equal performance or cost for every workload.

The strongest measured improvement over Kimi K3 was in agentic tasks: Artificial Analysis reported a 67% score on Terminal-Bench 4.0, compared with 13% for K3. On its GDPval-AA knowledge-work evaluation, K3.1 Pro scored 1,846 Elo, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, it scored 52%, below Opus 5.5 at 61%. These are benchmark results, not evidence that the model will perform the same way in every practical setting.

A Narrower Gap for Open Weights

The score places K3.1 Pro close to the leading closed models in this evaluation, extending competition beyond systems that users can access only through a provider’s service. If Moonshot releases usable weights on schedule, developers and organisations may be able to inspect, adapt or host the model themselves, subject to the eventual licence terms and hardware demands.

For now, the distinction between a model being available through an API and being available as open weights matters. Customers can use the hosted service, but cannot yet download K3.1 Pro’s weights to operate it independently. And although the listed API prices are competitive with some frontier offerings, the model’s 2.8-trillion total parameters mean that self-hosting could require substantial infrastructure. The release could widen deployment choices, but it does not make local operation straightforward or cost-free.

Benchmark proximity also needs careful interpretation. Artificial Analysis provides an independent comparison across its own tests, while a score does not establish that two models are interchangeable for a particular company’s coding, research or office workflow. The results give buyers evidence to test, rather than a substitute for testing on their own tasks.

What the Benchmarks Measure

The launch comes after a visible gap between the highest-scoring proprietary systems and available open-weight models on Artificial Analysis’s index. The source says the strongest open-weight model listed the day before K3.1 Pro’s launch was Xiaomi’s MiMo-V2.6-Pro, at 46.3, while Opus 5.5 scored 57.6. K3.1 Pro’s 55.1 score is a substantial step up on that comparison, though rankings depend on the benchmark, model configuration and date of measurement.

Moonshot has also presented demonstrations involving game development, 3D modelling and finance. These include claims that the model can build projects using Unity, Unreal Engine and Godot, and generate parametric models through Blender and CAD tools. The examples and many of the supporting tests are Moonshot’s own demonstrations or internal evaluations, rather than independent findings.

One example concerns hallucinations. Moonshot says an internal stress test measured a claim-level hallucination rate of 8.3% for K3.1 Pro, down from 23.4% for K3, describing that as a reduction of about 65%. Artificial Analysis’s separate AA-Omniscience evaluation recorded a smaller change, from 53% for K3 to 44% for K3.1 Pro. The tests use different methods, and Moonshot itself says its deliberately error-provoking test does not represent overall error rates in everyday use.

“The 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 key open question is whether Moonshot will publish the weights by its stated November 4 target. As of October 24, the source says they remained unavailable, and no K3.1 Pro licence had been published. Without the licence, users cannot determine what forms of commercial use, modification or redistribution will be allowed.

It is also unclear what hardware and operating costs would be required for organisations seeking to host the full model. The reported parameter counts signal a large system, but the available material does not specify a practical self-hosting setup or provide a complete cost comparison with hosted access. The benchmark results do not settle how the model performs across different production workloads, either.

Moonshot’s demonstrations and internal claims—including performance in games, CAD and finance—remain distinct from independently reproduced findings. Artificial Analysis’s evaluations offer external measurements for several tasks, but they do not verify every launch claim. The differing hallucination results underline that test design and measurement method affect the reported outcome.

The November Weights Deadline

The next milestone is Moonshot’s promised November 4 weights release. Publication, if it happens, should clarify the file availability and any accompanying licence terms; users will then be able to assess whether the release supports their intended use. Until those details appear, K3.1 Pro is accessible through Moonshot’s services rather than as a downloadable model.

Further independent evaluations and practical deployment reports will help show whether the launch-day benchmark lead translates into reliable performance outside test suites. Buyers can compare the hosted API against their own tasks and pricing needs, while organisations considering self-hosting will need to wait for technical and licensing details. The current evidence supports a near-frontier benchmark result; it does not yet establish an open, independently deployable model.

Key Questions

Is Kimi K3.1 Pro open-weight now?

No. It is available through Moonshot’s products and API, but the weights had not been released as of October 24. Moonshot said it plans to publish them by November 4.

How did K3.1 Pro rank?

Artificial Analysis gave it an Intelligence Index score of 55.1, ranking it fourth among the models listed in its October 24 figures. It trailed Opus 5.5, Sonnet 5.5 and Muse Wonder Full.

What does Moonshot charge for API use?

The source lists prices of $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimated about $2.11 per Intelligence Index task under its evaluation method.

Has Moonshot published a licence for the model?

No licence for K3.1 Pro had been published in the source material. Users will need the licence terms to determine what downloading, commercial use, modification and redistribution will be permitted.

Are the hallucination claims independently confirmed?

Not in the same form as Moonshot’s internal result. Moonshot reported an 8.3% rate on its own stress test, down from 23.4% for K3. Artificial Analysis measured a drop from 53% to 44% on a separate evaluation, which uses a different method.

Source: ThorstenMeyerAI.com