Kimi K3.1 Pro’s Frontier Chase Comes Before The Weights Release

🔍 Read the full analysis: Kimi K3.1 Pro’s Frontier Chase Comes Before The Weights Release on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get everyday essentials 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 released Kimi K3.1 Pro in its apps, Kimi Work, Kimi Code and API on Oct. 23, with the model maker describing it as its most capable 3T-class open-weight model. Artificial Analysis scored it 55.1 on its Intelligence Index, ranking it fourth among the models listed in the report; Moonshot says the weights will be published by Nov. 4, but has not yet disclosed a K3.1 licence.

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. Its 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. Those scores were recorded at each model’s highest reasoning setting.

The evaluation shows a marked change from the earlier Kimi K3 on some tasks. K3.1 Pro reached 67% on Artificial Analysis’s Terminal-Bench 4.0 run, compared with 13% for K3. It also scored 1,846 Elo on GDPval-AA, a simulated knowledge-work evaluation, below Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, K3.1 Pro scored 52%, compared with 47% for K3 and 61% for Opus 5.5.

Artificial Analysis lists K3.1 Pro at $2 per million input tokens and $10 per million output tokens, estimating a cost of $2.11 per Intelligence Index task. The corresponding estimated task cost for Opus 5.5 is $5.98. These figures compare costs under the evaluation’s pricing and task assumptions; they are not a guarantee of what every customer will pay for a particular workload.

A Narrower Gap at Lower Test Cost

The reported scores put K3.1 Pro close to leading proprietary models on this particular benchmark, while its listed API rates are lower than Opus 5.5’s. For developers and businesses, that combination could make it worth testing for coding and agentic workflows, where the model’s Terminal-Bench result is particularly strong. Benchmark scores, however, do not establish performance across every production task, nor do they show the full cost of deploying a model.

The possible open-weight release is a separate part of the story. Artificial Analysis describes K3.1 Pro as likely to become the leading open-weight model once its weights are published. That wording is conditional: the model is currently available through Moonshot’s services, while the report says its weights have not yet been released. A downloadable model could give organizations more control over deployment and data handling, subject to the licence and hardware requirements.

Moonshot’s claims about end-to-end game creation, Blender and CAD modelling, and fewer hallucinations may attract users, but they need to be read as vendor claims unless independently reproduced. The distinction matters for procurement decisions: a polished demo or internal test can identify capabilities worth evaluating, but it is not equivalent to evidence of reliable performance in a company’s own environment.

From Kimi K3 to K3.1 Pro

K3.1 Pro is described by Moonshot as a more heavily post-trained version of Kimi K3. On the Artificial Analysis figures supplied in the report, K3 scored 43.6 on the Intelligence Index, against 55.1 for K3.1 Pro. The comparison suggests a substantial improvement on that index, but it does not isolate which post-training changes produced the difference.

Before this launch, the report lists Xiaomi’s MiMo-V2.6-Pro at 46.3 as the highest-scoring open-weight model in the comparison, while Opus 5.5 scored 57.6. K3.1 Pro’s score narrows that benchmark gap. But it is not yet an open-weight release: Artificial Analysis lists its weights as promised by Nov. 4, and the source says Moonshot had not published the K3.1 licence as of Oct. 24.

Moonshot’s launch materials also describe internal tests and demonstrations. The company reported a 65% reduction in claim-level hallucinations on its Kimi Hallu Bench Hard, from 23.4% for K3 to 8.3% for K3.1 Pro. The source notes that Moonshot calls the test deliberately designed to provoke errors and says its results do not represent overall error rates in daily use. On Artificial Analysis’s separate AA-Omniscience measure, the reported rate fell from 53% for K3 to 44% for K3.1 Pro.

“Its most capable “3T-class open-weight model.””

— Moonshot AI, describing Kimi K3.1 Pro

Weights, Licence and Real-World Results

The weights are not yet available, and the Nov. 4 date is a promise reported by the source rather than a completed release. The K3.1 licence has not been published, so prospective users cannot yet assess its terms for commercial use, modification or redistribution. Those details could affect whether organizations can use the model as they hope.

There are also practical limits to self-hosting. The model’s listed scale is 2.8 trillion total parameters, and the source says running it independently would require data-centre-level resources. The actual hardware configuration, deployment options and operating costs are not supplied. It is also unclear whether the benchmark ranking will hold across other evaluations or users’ production workloads.

Moonshot’s demonstrations and internal measurements have not all been independently verified in the source material. Artificial Analysis provides separate benchmark results, including a smaller decline in hallucination rate than Moonshot’s internal test reports. The two figures measure different evaluations and should not be treated as direct contradictions or interchangeable estimates.

The Nov. 4 Weights Deadline

The next test is whether Moonshot publishes the K3.1 Pro weights by its stated Nov. 4 deadline, and what licence accompanies them. Users and organizations will then be able to evaluate the release terms and, where their hardware permits, test deployment outside Moonshot’s hosted products.

Until then, K3.1 Pro can be accessed through the company’s apps, Kimi Work, Kimi Code and API, at the prices listed by Artificial Analysis. Further independent testing across coding, knowledge work and hallucination evaluations would help clarify how well the launch-day results translate to day-to-day use. The source does not specify a separate date for additional evaluations or a further Moonshot announcement.

Key Questions

When did Kimi K3.1 Pro launch?

Moonshot AI released it on Oct. 23, 2026, in the Kimi apps, Kimi Work, Kimi Code and the Kimi API.

How did Kimi K3.1 Pro score?

Artificial Analysis gave it 55.1 on its Intelligence Index, placing it fourth in the model comparison reported on Oct. 24. The same evaluation recorded 67% on Terminal-Bench 4.0.

Are Kimi K3.1 Pro’s weights available now?

No. The source says Moonshot promised the weights by Nov. 4. As of Oct. 24, they had not been published, and a K3.1 licence was not available.

What does Moonshot claim about hallucinations?

Moonshot reported that its internal Kimi Hallu Bench Hard claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro. The company cautioned that the test is deliberately designed to provoke errors and does not represent overall daily-use error rates. Artificial Analysis reported a decline from 53% to 44% on its separate AA-Omniscience evaluation.

Can companies run the model themselves?

That cannot yet be established from the available details. The weights and licence have not been published, and the model’s 2.8-trillion total parameter count points to substantial infrastructure needs. The source characterizes self-hosting at this scale as a data-centre project.

Source: ThorstenMeyerAI.com