Kimi K3.1 Pro Nears The Frontier—Its Weights Are Still Unreleased

🔍 Read the full analysis: Kimi K3.1 Pro Nears The Frontier—Its Weights Are Still Unreleased on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets 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 launched Kimi K3.1 Pro in its apps, Kimi Work, Kimi Code and API on Oct. 23, with an independent benchmark placing it fourth overall behind three proprietary models. The company says it will release the model weights by Nov. 4, but as of Oct. 24 they were not available and no licence had been published, limiting what users can verify or run independently.

Artificial Analysis gave K3.1 Pro an Intelligence Index score of 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 evaluation compares models at their highest reasoning settings. Artificial Analysis said K3.1 Pro would be the leading open-weights model once its weights are published; that status is conditional, not a current availability claim.

The evaluation lists K3.1 Pro as a 2.8-trillion-parameter mixture-of-experts model, with 104 billion active parameters, a one-million-token context window and text and image input. Artificial Analysis estimated a cost of about $2.11 per Intelligence Index task, compared with $5.98 for Opus 5.5. The listed API rates are $2 per million input tokens and $10 per million output tokens. Such task-cost estimates reflect the evaluation’s setup and are not a guarantee of costs for every workload.

Artificial Analysis recorded a 67% Terminal-Bench 4.0 score, compared with 13% for the earlier Kimi K3, and 1,846 Elo on its GDPval-AA knowledge-work evaluation. It also measured a fall in claim-level hallucination rate on its AA-Omniscience test from 53% for K3 to 44% for K3.1 Pro. These are results from specific benchmarks, not general measures of performance in all uses.

A Narrower Gap to Closed Models

The score places K3.1 Pro close to the highest-ranked proprietary systems in Artificial Analysis’s comparison, while the listed API price is lower than Opus 5.5’s. That combination could matter to organizations choosing models for coding, research or agentic tasks, particularly if performance holds across their own workloads. Benchmark rankings and token prices alone, however, do not establish which model will be cheaper or more reliable for a particular job.

The larger potential change depends on the weights becoming available under terms that permit meaningful use. Open weights can allow deployment and adaptation outside a provider’s hosted service, but access is not the same as unrestricted use: the licence controls what users may do, and running a model listed at 2.8 trillion total parameters would require substantial infrastructure. Until both weights and licence are published, K3.1 Pro’s open-model status remains prospective.

What the Benchmarks Actually Show

Moonshot describes K3.1 Pro as a more heavily post-trained version of Kimi K3 and says it matches leading proprietary models on coding, agentic and knowledge benchmarks. The independent comparison offers a mixed picture: the sharpest reported gain over K3 is on Terminal-Bench, while Humanity’s Last Exam moved from 47% to 52%; Opus 5.5 scored 61% on that test. A strong result on one task category should not be read as a top ranking across every capability.

Moonshot also presented demonstrations involving game development in Unity, Unreal Engine and Godot, parametric modeling in Blender and CAD, and finance work. The company reported a 92.7% BenchCAD score with tools, below its cited scores for Opus 5.5 and GPT-6 Astra. For hallucinations, Moonshot’s internal Kimi Hallu Bench Hard showed a decline from 23.4% for K3 to 8.3% for K3.1 Pro. The company described that test as deliberately designed to provoke errors and said its rates do not represent everyday error rates.

Artificial Analysis’s separate AA-Omniscience evaluation found a smaller decline, from 53% to 44%. The different results reflect different tests and methods; they are not directly interchangeable. Moonshot’s demonstrations and internal benchmark claims have not been independently verified in the supplied material.

“The Kimi Hallu Bench Hard test is “deliberately crafted” to provoke errors, and its rates “do not represent overall error rates in daily use.””

— Moonshot AI

Weights, Licence and Real-World Results

As of Oct. 24, the weights had not been released, despite Moonshot’s stated target of Nov. 4, and no K3.1 Pro licence had been published. The available information does not establish whether the deadline will be met, what restrictions the licence may impose, or whether the model will be downloadable in a form practical for particular users.

There is also no independent confirmation here of Moonshot’s game-development, CAD and finance demonstrations. Benchmark outcomes can vary with prompts, tools, reasoning settings and evaluation design. The reported hallucination figures come from distinct tests, and neither the internal test nor AA-Omniscience establishes how often the model will make errors in ordinary use.

Infrastructure needs and total ownership costs remain uncertain for self-hosting. The published API rates provide one pricing reference, but they do not show the cost of running the model locally or the expense of serving it at scale.

The Nov. 4 Release Test

The next stated milestone is Nov. 4, 2026, when Moonshot says it plans to publish K3.1 Pro’s weights. A release would allow users and researchers to inspect the files, test performance outside the company’s hosted products and assess whether the model is usable on their infrastructure. The accompanying licence will be key to determining what deployment, modification and commercial uses are allowed.

Until that release, K3.1 Pro can be assessed through its available Moonshot products and the published benchmark results, but not as a model that users can independently download and run. Whether the weights arrive on schedule, and whether independent testing reproduces the reported results, remain open questions.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a Moonshot AI model released in the company’s apps, Kimi Work, Kimi Code and API 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 did it rank in independent testing?

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

Can users download the weights now?

No. As of Oct. 24, Moonshot had not released the weights. The company said it plans to publish them by Nov. 4, but that date is a stated target, not a completed release.

What licence will govern the weights?

No licence for K3.1 Pro had been published as of Oct. 24. Without it, users cannot confirm which forms of deployment, modification or commercial use will be permitted.

Are Moonshot’s hallucination claims independently confirmed?

Not on the same test. Moonshot reported an internal reduction from 23.4% to 8.3% on Kimi Hallu Bench Hard. Artificial Analysis measured a decline from 53% to 44% on its separate AA-Omniscience evaluation, which is a smaller change on a different benchmark.

Source: ThorstenMeyerAI.com