Kimi K3.1 Pro Is Near The Frontier, But Access To Its Weights Has To Wait

🔍 Read the full analysis: Kimi K3.1 Pro Is Near The Frontier, But Access To Its Weights Has To Wait on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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 released Kimi K3.1 Pro through its apps, coding product and API on October 23, while keeping the model weights unavailable for now. Artificial Analysis gave it a 55.1 Intelligence Index score, fourth among models in its comparison, but Moonshot’s promise to publish the weights by November 4 is still conditional on a release that has not happened.

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 a one-million-token context window and text and image inputs. The service is available through Kimi apps, Kimi Work, Kimi Code and the Kimi API. Moonshot’s published API prices are $2 per million input tokens and $10 per million output tokens.

In Artificial Analysis’s comparison, K3.1 Pro scored 55.1, behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The evaluator estimated a cost of $2.11 per Intelligence Index task for K3.1 Pro, compared with $5.98 for Opus 5.5 at its maximum reasoning setting. These are evaluator-specific estimates, not a guarantee of what a particular user or workload will pay.

The strongest reported result was on Terminal-Bench 4.0, where Artificial Analysis measured a 67% score, compared with 13% for the earlier Kimi K3. On GDPval-AA, its simulated knowledge-work evaluation, K3.1 Pro scored 1,846 Elo, below Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, it scored 52%, while Opus 5.5 scored 61%.

Moonshot has also promoted demonstrations involving game engines, parametric 3D design and financial work. Those examples and many associated performance claims come from the company’s own tests. Moonshot reports that its internal Kimi Hallu Bench Hard claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro. Artificial Analysis measured a smaller change on its separate AA-Omniscience evaluation: from 53% for K3 to 44% for K3.1 Pro.

Why the Missing Weights Matter

The score puts K3.1 Pro close to top proprietary models on one independent benchmark suite, while its measured coding-agent result marks a substantial improvement over the previous Kimi model on Terminal-Bench 4.0. For developers and businesses, that may make the model worth testing through Moonshot’s hosted products, particularly for tasks involving tool use and longer workflows.

But a strong score does not make the model open to inspection or self-hosting. Without published weights and a licence, organizations cannot yet verify that they can run it on their own infrastructure or use it under terms that suit their needs. That distinction matters for teams with data-control requirements, for researchers who want to reproduce results, and for developers planning around the cost and availability of a hosted API.

Even if the weights arrive, the model’s scale will limit who can run it. A system listed at 2.8 trillion total parameters is a substantial infrastructure undertaking, not a routine download for an individual laptop. The practical value of a release will depend on the licence, formats, hardware requirements and any deployment options Moonshot provides.

From Kimi K3 to K3.1 Pro

Kimi K3.1 Pro is described in the supplied reporting as a more heavily post-trained version of Kimi K3. Artificial Analysis scored Kimi K3 at 43.6 in its comparison; K3.1 Pro scored 55.1. The evaluator’s results therefore show a notable difference between the two versions in that index, though the benchmark score alone does not establish how the models perform across every real-world task.

Before K3.1 Pro’s launch, Artificial Analysis listed Xiaomi’s MiMo-V2.6-Pro at 46.3 as the highest-scoring open-weight model in the supplied comparison. K3.1 Pro’s 55.1 score is higher, but the evaluator described it as set to become the leading open-weight model once its weights are published. At present, the rankings include its performance while the weights remain unavailable, so its open-weight status is prospective rather than current.

Moonshot’s internal tests add examples beyond the independent scorecard, including a simulated walking project and reported improvements in financial modelling and research. Those claims may help explain the company’s intended use cases, but they should be distinguished from third-party measurements. Moonshot itself says its hallucination test was deliberately crafted to provoke errors and that its rates do not represent overall error rates in everyday use.

“The weights are promised “by Nov 4.””

— Moonshot AI, on the reported weights release

Weights, Licence and Claims Pending

As of the October 24 report, K3.1 Pro’s weights had not been published, and no licence for the model had been released. It is therefore not yet clear whether Moonshot will meet its stated November 4 date, what use or redistribution terms it will set, or what hardware and software formats will be supported.

The independent benchmark results are measurements from Artificial Analysis, not proof that the model will match those scores in every deployment. Moonshot’s demonstrations of full game projects, CAD work and finance tasks are company claims based largely on its own testing. The supplied material does not establish how consistently those results can be reproduced by outside users or how they compare across a broad range of everyday tasks.

The benchmark figures also have limits. The 65% reduction in hallucinations comes from Moonshot’s internal comparison of 23.4% and 8.3% on Kimi Hallu Bench Hard; Artificial Analysis reported a different result on a separate evaluation. These measures should not be treated as interchangeable, and the available information does not establish an overall error rate for ordinary use.

The November 4 Release Test

The next milestone is Moonshot’s promised publication of the K3.1 Pro weights by November 4. A release would let users examine the files and determine whether the model can be deployed under the terms Moonshot provides. Until then, access remains through the company’s hosted products and API, rather than a public weights download.

Readers evaluating the model should watch for the licence text, release formats, hardware guidance and any independent testing after publication. Those details will help establish whether the model is practically usable outside Moonshot’s services—not just how it ranks in a benchmark. The supplied information does not confirm a release beyond the company’s stated date.

Key Questions

When did Kimi K3.1 Pro launch?

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

How did Kimi K3.1 Pro rank?

Artificial Analysis scored it 55.1 on its Intelligence Index, placing it fourth in the supplied comparison, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.

Can users download the model weights now?

No. The report says the weights were not yet published as of October 24. Moonshot said it planned to release them by November 4.

Has Moonshot published a licence for K3.1 Pro?

No K3.1 Pro licence had been published in the supplied reporting. Without it, the terms for modifying, redistributing or commercially using released weights remain unknown.

Are the hallucination claims independently confirmed?

Not in the same form as Moonshot’s claim. Moonshot reported a fall from 23.4% to 8.3% on its internal Kimi Hallu Bench Hard test; Artificial Analysis measured a fall from 53% to 44% on a separate evaluation. The tests use different measures, and Moonshot cautioned that its internal test does not represent everyday error rates.

Source: ThorstenMeyerAI.com