Three Points From The Frontier: Meet Kimi K3.1 Pro

🔍 Read the full analysis: Three Points From The Frontier: Meet Kimi K3.1 Pro 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 released Kimi K3.1 Pro on October 23 in its Kimi apps, Kimi Work, Kimi Code and API, presenting it as a more heavily post-trained version of Kimi K3. Artificial Analysis scored it 55.1 on its Intelligence Index, ranking it fourth in the comparison, but the model is not yet downloadable: Moonshot has promised to publish its weights by November 4, and has not published 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 Intelligence Index score trails Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. These scores reflect the models at their highest reasoning settings in the figures reported on October 24.

The evaluation suggests a particularly large improvement in agentic tasks over Kimi K3. Artificial Analysis measured 67% on Terminal-Bench 4.0 for K3.1 Pro, compared with 13% for K3. The same source reported a score of 1,846 Elo on GDPval-AA, a simulated knowledge-work evaluation, behind 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 API pricing at $2 per million input tokens and $10 per million output tokens. It estimates a cost of $2.11 per Intelligence Index task, compared with $5.98 for Opus 5.5 at maximum effort. That is a benchmark-specific cost estimate, not a guarantee that every real-world task will cost the same. The source also measured the K3.1 Pro hallucination rate at 44% on its AA-Omniscience evaluation, down from 53% for K3.

A Smaller Gap to Closed Models

K3.1 Pro’s score places it close to the top models in this evaluation, and its measured performance on Terminal-Bench stands out against Kimi K3. If the weights are released as promised, developers and organisations could inspect and run a model with a score near leading proprietary systems rather than accessing it only through a vendor’s hosted product. That could matter for teams seeking greater control over deployment, data handling or customisation.

That prospect is conditional, not a present capability. The weights are still pending and the licence is unpublished, so users cannot yet confirm the terms for commercial use, modification or redistribution. Nor does a benchmark ranking establish how a model will perform across a company’s actual workloads. The listed total parameter count also means self-hosting could require substantial infrastructure; the source characterises it as a data-centre-scale undertaking.

The price comparison may also attract API customers, particularly for workloads similar to the benchmark tasks. But the published task-cost estimate is tied to Artificial Analysis’s evaluation method. Actual bills will depend on token use, reasoning settings, tool calls and the task being run.

From Kimi K3 to K3.1 Pro

Moonshot describes K3.1 Pro as a more heavily post-trained version of Kimi K3. The earlier model scored 43.6 on Artificial Analysis’s Intelligence Index, while K3.1 Pro scored 55.1 in the reported comparison. The largest cited change is on Terminal-Bench 4.0, where the measured score rose from 13% to 67%; the gains on Humanity’s Last Exam were smaller, rising from 47% to 52%.

Moonshot has also publicised demonstrations involving game engines, 3D modelling and financial work. The company says the model can create projects in Unity, Unreal Engine and Godot, and generate parametric 3D work using Blender and CAD software. It reported a 92.7% result on BenchCAD with tools, below the figures it cited for Claude Opus 5.5 and GPT-6 Astra. These examples and several financial-work results come from Moonshot’s own demonstrations or internal tests; they should not be confused with the separate Artificial Analysis evaluations.

The company’s internal Kimi Hallu Bench Hard test reported a claim-level hallucination rate falling from 23.4% for K3 to 8.3% for K3.1 Pro. Moonshot describes that as a roughly 65% reduction, while also warning that the test was deliberately designed to provoke errors and does not represent error rates in everyday use. Artificial Analysis’s separate AA-Omniscience evaluation showed a smaller change, from 53% to 44%.

“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

The most consequential unresolved detail is whether Moonshot will publish K3.1 Pro’s weights by its stated November 4 deadline. The company has promised a release, but the weights were not available in the source material dated October 24. The model’s K3.1 licence also had not been published, leaving usage and redistribution terms unclear.

Some performance claims remain difficult to assess independently. Moonshot’s game-building demonstrations, internal financial evaluations and Kimi Hallu Bench Hard results are vendor-reported; the supplied material does not provide independent replication of those specific tests. The benchmarks that Artificial Analysis did run offer a separate measure, but they cannot settle how K3.1 Pro will behave across different products, prompts and operating conditions.

It is also unclear what hardware and operating costs would be required to run the complete model outside Moonshot’s services. Its total parameter count is 2.8 trillion, but the material does not give a practical deployment configuration or verified self-hosting cost. No conclusion about everyday error rates or company-wide savings follows directly from the benchmark results.

The November 4 Release Test

The next public milestone is Moonshot’s promised November 4 weights release. Readers and developers will be able to assess whether the files appear on schedule and review the accompanying licence, if both are published. Those details will determine whether K3.1 Pro can fairly be treated as an open-weight option and who may use or adapt it.

After release, independent testing and reports from developers could show whether the benchmark results hold up in different settings and what infrastructure is needed to run the model. Until then, K3.1 Pro is available through Moonshot’s listed products and API, while its open-weight status remains a future claim rather than a completed release.

Key Questions

When did Kimi K3.1 Pro launch?

Moonshot AI made it available 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 gave it a 55.1 Intelligence Index score and ranked it fourth in the comparison reported on October 24, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.

Can users download the model weights now?

Not according to the source material dated October 24. Moonshot promised to publish the weights by November 4, but the release had not yet occurred.

What is known about the K3.1 licence?

No K3.1 licence had been published by October 24. Until terms are available, permissions for commercial use, modification and redistribution remain unclear.

Are Moonshot’s hallucination claims independently verified?

The reported 65% reduction comes from Moonshot’s internal Kimi Hallu Bench Hard test, which the company says is deliberately designed to provoke errors. Artificial Analysis reported a smaller decline on its separate AA-Omniscience evaluation, from 53% for K3 to 44% for K3.1 Pro.

Source: ThorstenMeyerAI.com