Kimi K3.1 Pro Comes Within Three Points As Its Weights Remain Unreleased

🔍 Read the full analysis: Kimi K3.1 Pro Comes Within Three Points As Its Weights Remain Unreleased on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get privacy and security gear 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’s Kimi K3.1 Pro scored 55.1 on Artificial Analysis’s Intelligence Index, placing fourth in the service’s October 24 comparison and three points behind its top-ranked model. The model is available through Moonshot’s products and API, but its weights had not been published, and the company had not released a K3.1 licence—two conditions that matter to developers considering whether they can run and use it independently.

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. Moonshot describes it as a more heavily post-trained version of Kimi K3. The company made it available on October 23 through the Kimi apps, Kimi Work, Kimi Code and its API.

In the Artificial Analysis ranking, K3.1 Pro’s 55.1 score sits below Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The table compares models at their highest listed reasoning settings. Artificial Analysis describes K3.1 Pro as a prospective leader among open-weight models once its weights are published; at the time of its assessment, it was not downloadable as an open-weight model.

Artificial Analysis lists K3.1 Pro’s cost at $2 per million input tokens and $10 per million output tokens, and estimates $2.11 per Intelligence Index task. It estimates $5.98 per task for Claude Opus 5.5 at maximum effort. These are evaluation-based cost figures, not a guarantee of what every user or workload will pay.

A Narrower Gap, With Conditions

If Moonshot publishes the weights as promised, the result could give developers access to a model whose measured score is close to the leading closed systems. Artificial Analysis’s comparison puts K3.1 Pro three points below its top entry, while the best open-weight model in the comparison before this launch, Xiaomi’s MiMo-V2.6-Pro, scored 46.3. That makes the reported result relevant to organizations weighing model capability against the control and deployment options that downloadable weights can offer.

But the score alone does not establish that K3.1 Pro is ready for local or private deployment. The weights are still pending, and the licence is unknown. Further, Artificial Analysis lists 2.8 trillion total parameters. That scale suggests running the model directly could require substantial infrastructure; the source article characterizes it as a data-centre project. The cost advantage reported in the benchmark also depends on the stated API prices and evaluation workload, rather than proving lower costs for every application.

The results are strongest in some areas and less decisive in others. Artificial Analysis reports a 67% Terminal-Bench 4.0 score, compared with 13% for Kimi K3, and 1,846 Elo on GDPval-AA, behind two models in its comparison. On Humanity’s Last Exam, K3.1 Pro scored 52%, while Opus 5.5 scored 61%. Those differences make it more useful to treat the model as a notable benchmark result than as a blanket replacement for proprietary systems.

What the Benchmarks Measure

The launch combines an independent evaluation with company-reported demonstrations. Artificial Analysis says it ran its own tests on launch day. Its results show a large increase over Kimi K3 on Terminal-Bench 4.0, but smaller gains on some knowledge measures: Humanity’s Last Exam rose from 47% for K3 to 52% for K3.1 Pro. On Artificial Analysis’s AA-Omniscience measure, the claim-level hallucination rate fell from 53% to 44%.

Moonshot separately advertises coding, agentic work, knowledge tasks and software-assisted 3D design. The company reports a 92.7% result on BenchCAD with tools, below the figures it gives for Claude Opus 5.5 and GPT-6 Astra. It also reports an OfficeQA Pro v2 score of 64.4 and internal gains in financial modelling and research compared with Kimi K3. These are company-provided results, and they should not be confused with the independent Intelligence Index ranking.

The distinction matters for the most widely repeated hallucination claim. Moonshot says K3.1 Pro’s claim-level hallucination rate fell from 23.4% to 8.3% on its internal Kimi Hallu Bench Hard test, describing that as about a 65% reduction. The company also says that test was deliberately designed to provoke errors and that the rates do not represent overall error levels in daily use. Artificial Analysis’s separate AA-Omniscience result shows a decline from 53% to 44%, a different measure and a smaller change.

“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 and Licence Still Pending

The release status is the main unresolved issue. Moonshot had promised the weights by November 4, but they were not available at the time of the October 24 report. It is not clear whether the company will meet that date, what form the release will take, or whether access will come with restrictions. No K3.1 licence had been published, so prospective users could not yet assess the terms for modification, redistribution or commercial use.

There are also limits to what the benchmark results establish. Artificial Analysis provides an independent comparison, but the supplied material does not detail every test condition or show how scores translate to a particular company’s workloads. Moonshot’s examples of end-to-end game creation, finance work and 3D modelling are vendor demonstrations; their performance has not been independently confirmed in the source material. The precise practical cost of operating the model directly is also not established.

Watch for the November Release

The next milestone is Moonshot’s promised November 4 weight release. If it happens, the release materials should clarify the licence, available model files and any requirements for deployment. Those details will determine whether developers can inspect, adapt or run K3.1 Pro under terms that fit their projects.

Until then, the model can be assessed through Moonshot’s hosted products and API, while the benchmark result remains a measurement of a model whose weights are not yet public. Readers should distinguish that access from a downloadable release and treat Moonshot’s internally tested demonstrations separately from Artificial Analysis’s evaluation. Any change to the release date, licence or evaluation evidence remains to be confirmed.

Key Questions

What did Kimi K3.1 Pro score?

Artificial Analysis gave it an Intelligence Index score of 55.1 in its October 24, 2026 comparison, placing it fourth among the listed models.

Are Kimi K3.1 Pro’s weights available?

Not as of October 24, 2026. Moonshot said it planned to publish the weights by November 4.

Has Moonshot published a licence for K3.1?

No K3.1 licence had been published in the source material. Until the terms are available, users cannot confirm the conditions for modifying, redistributing or commercially using the weights.

Are the hallucination figures independently verified?

Moonshot’s claim of about a 65% reduction comes from its internal Kimi Hallu Bench Hard test. Artificial Analysis reported a separate reduction on AA-Omniscience, from 53% to 44%; the two results use different tests.

Can the benchmark score predict how K3.1 Pro will perform for my work?

Not by itself. The index summarizes performance across evaluations, while results on a particular task can differ. The source material does not establish how the model will perform across individual users’ workloads.

Source: ThorstenMeyerAI.com