Kimi K3.1 Pro Scores Near The Frontier As Weights Remain Unreleased

🔍 Read the full analysis: Kimi K3.1 Pro Scores Near The Frontier As Weights Remain Unreleased 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 an evaluation by Artificial Analysis placed it fourth on the organisation’s Intelligence Index, at 55.1 points. The result puts the model three points behind the index leader, but it is not yet an open-weights release: Moonshot says the weights will be published by Nov. 4, and no licence for K3.1 has been made public.

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 1-million-token context window and text and image input. Moonshot has made it available through the Kimi apps, Kimi Work, Kimi Code and its API. Artificial Analysis’s reported score is 55.1, compared with 57.6 for Claude Opus 5.5 at maximum reasoning, 56.0 for Claude Sonnet 5.5 and 55.8 for Muse Wonder Full.

The same evaluation puts K3.1 Pro at $2.11 per Intelligence Index task, compared with $5.98 for Opus 5.5 at maximum effort. The listed API prices are $2 per million input tokens and $10 per million output tokens. Artificial Analysis also recorded 67% on Terminal-Bench 4.0, a benchmark of agentic coding tasks. That is higher than the results listed for the other models in the source table, including 60% for Opus 5.5, though benchmark results do not by themselves establish performance across everyday uses.

Moonshot’s own launch materials describe demonstrations involving game engines, 3D modelling, CAD and financial work. The company reports a 65% reduction in claim-level hallucinations on its internal Kimi Hallu Bench Hard, from 23.4% for K3 to 8.3% for K3.1 Pro. Artificial Analysis measured a smaller change on its separate AA-Omniscience evaluation: the reported hallucination rate fell from 53% for K3 to 44% for K3.1 Pro. The figures come from different tests and should not be treated as directly comparable.

A Near-Frontier Score, Not Yet Open

The benchmark result matters because it places a model that Moonshot intends to release with weights close to the leading proprietary systems on one independent ranking. If the weights are published, researchers and organisations could inspect, adapt or run the model under whatever terms its eventual licence permits. That could offer an alternative to relying solely on hosted services from closed-model providers.

Those possibilities are not available yet. The weights are still pending, the licence is unknown, and the listed model is a very large system. Its total size of 2.8 trillion parameters means local deployment may require substantial infrastructure; the source describes self-hosting it as a data-centre-scale undertaking. The advertised lower cost per benchmark task is relevant to API customers, but it does not establish the full cost of operating the model independently.

The gap between vendor demonstrations and independent measurements also matters for buyers. Moonshot’s examples suggest possible strengths in longer coding workflows and specialist tasks, while the external benchmark offers a separate but limited snapshot. Companies evaluating the model still need to test it on their own workloads, costs, data rules and reliability requirements.

How K3.1 Compares With K3

K3.1 Pro is described by the source as a more heavily post-trained version of Moonshot’s Kimi K3. In Artificial Analysis’s figures, the earlier K3 scored 43.6 on the Intelligence Index, while K3.1 Pro scored 55.1. The listed Terminal-Bench 4.0 result rose from 13% to 67%. These are results from the evaluation cited in the source, not a guarantee that every user or task will see a similar improvement.

The reported index score also narrows the gap between the strongest proprietary model in the table and the model that may become an open-weights release. Artificial Analysis’s description is conditional: it said K3.1 Pro “will be the leading open-weights model once its weights are published.” As of Oct. 24, the weights had not been released, so the distinction between a model accessible through Moonshot’s services and one available for independent deployment remains material.

Moonshot’s launch material includes a range of company-reported results, including a score of 92.7% on BenchCAD with tools and 64.4 on OfficeQA Pro v2. The source identifies these as vendor claims or internal results. They should be read separately from Artificial Analysis’s evaluation, which tested the model using its own benchmarks and published comparative figures.

“The company says it will publish the weights “by Nov 4.””

— Moonshot AI

Weights, Licence and Real-World Results

The central open question is whether Moonshot will publish the weights by its promised Nov. 4 date. The source says they had not been released by Oct. 24. It also reports that no K3.1 licence had been published, leaving unclear what users would be permitted to do with the weights if they arrive. The model’s score and availability through Moonshot’s services do not resolve either issue.

It is also unclear how well the model’s benchmark performance will carry over to production tasks. Artificial Analysis provides independent measurements, but its scores represent specific test setups. Moonshot’s demos and internal evaluations provide additional evidence from the company, not independent confirmation. The source does not provide enough information to verify all the demonstrations or assess error rates across routine customer use.

Finally, the headline price comparison applies to the reported cost per Intelligence Index task, not every application. Actual API bills depend on usage, and self-hosting costs would depend on hardware, serving requirements and operational needs. The source does not provide a full cost analysis for private deployment.

The Nov. 4 Release Test

The next stated milestone is Moonshot’s promised publication of K3.1 Pro’s weights by Nov. 4. A release would allow outside users to examine the files and assess whether the model can be deployed independently. Any accompanying licence will determine the permitted forms of use, modification and redistribution, so publication of weights alone would not answer every access question.

Further independent testing could clarify how the model performs on coding, knowledge work and hallucination evaluations beyond the launch-day results. For now, buyers can access K3.1 Pro through Moonshot’s listed apps, work and coding products, and API, while organisations considering self-hosting must wait for the weights and licence details. No later release or licensing terms are confirmed in the source material.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a Moonshot AI model that the company describes as a more heavily post-trained version of Kimi K3. Artificial Analysis lists it as a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters.

How did it rank?

Artificial Analysis gave it a 55.1 Intelligence Index score and ranked it fourth among the models in the cited results. It was three points below Claude Opus 5.5 at maximum reasoning, which scored 57.6.

Can users download and run the weights now?

No. As of Oct. 24, 2026, the weights had not been published. Moonshot said it planned to release them by Nov. 4, but the source does not confirm that release has occurred.

What licence will apply to the weights?

No K3.1 licence had been published as of Oct. 24, according to the source. The terms governing use, modification and redistribution therefore remain unknown.

Are Moonshot’s performance claims independently verified?

Some comparative figures come from Artificial Analysis’s independent evaluation. Moonshot’s game, CAD, finance and internal hallucination results are company-reported claims; they are not the same as the external benchmark measurements.

Source: ThorstenMeyerAI.com