Near The Frontier, Still Without Weights: Kimi K3.1 Pro

🔍 Read the full analysis: Near The Frontier, Still Without Weights: Kimi K3.1 Pro on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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 on Oct. 23 through its Kimi apps, Kimi Work, Kimi Code and API, presenting it as a high-capability model that could approach leading proprietary systems. Artificial Analysis scored it 55.1 on its Intelligence Index, fourth among the models in its comparison; however, the weights have not been released, and the licence remains unpublished.

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 of 55.1 trails Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The score is based on the evaluated model and does not mean the weights are available to download or run independently.

The strongest result in Artificial Analysis’s reported task comparisons was 67% on Terminal-Bench 4.0, a measure of agentic coding performance. The outlet reported a score of 13% for the earlier Kimi K3 model on that benchmark. K3.1 Pro also scored 1,846 Elo on GDPval-AA, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, it scored 52%, compared with 47% for K3 and 61% for Opus 5.5.

Artificial Analysis listed API prices of $2 per million input tokens and $10 per million output tokens. It estimated an Intelligence Index task cost of $2.11 for K3.1 Pro, compared with $5.98 for Opus 5.5 at maximum reasoning. These are provider and evaluator figures for API use, not the costs of operating a local deployment. Moonshot’s planned release is by Nov. 4; until then, users access the model through the company’s services.

A Narrower Gap, With Access Still Central

The result matters because it places a model that Moonshot says will be open-weight close to the leading proprietary models on one independent aggregate benchmark. Artificial Analysis reported a 55.1 score for K3.1 Pro and 57.6 for Opus 5.5. That is a small numerical gap on this index, though it does not show that the models perform equally across all tasks or in everyday use.

For developers and organizations, the distinction between model performance and model access is central. A published set of weights could allow examination, modification and self-hosting, subject to the eventual licence and hardware requirements. But the model is not open for those uses yet. Its listed size—2.8 trillion total parameters—also makes self-hosting a substantial infrastructure undertaking, rather than a straightforward substitute for a hosted API.

The price comparison may be relevant to teams buying API access, but it is not a full cost comparison: workloads, reasoning settings and usage patterns affect spending. The benchmark results offer evidence of capability under defined tests, not a guarantee of lower costs or better results in a company’s own systems. The release date and licence will determine whether the promised openness changes practical access.

From Kimi K3 to K3.1 Pro

K3.1 Pro is described in the source material as a more heavily post-trained version of Kimi K3. The previous Kimi K3 scored 43.6 on the Artificial Analysis Intelligence Index in the comparison supplied, while K3.1 Pro scored 55.1. The same comparison listed Xiaomi’s MiMo-V2.6-Pro at 46.3 among open-weight models and Opus 5.5 at 57.6 among closed models.

Moonshot’s launch materials describe work on game development, parametric 3D modelling, CAD and finance. The company says the model created projects using Unity, Unreal Engine and Godot and reports a 92.7% result on BenchCAD with tools. These examples and figures are Moonshot’s own claims; they should not be treated as independently verified results. The source also reports an internal Kimi Hallu Bench Hard claim-level hallucination rate of 8.3%, down from 23.4% for K3. In contrast, Artificial Analysis measured a decline on its AA-Omniscience test from 53% to 44%.

The difference between those hallucination figures reflects distinct tests, not a single universal rate. Moonshot described its internal test as deliberately designed to provoke errors and cautioned that it did not represent overall error rates in daily use. Artificial Analysis’s result is also tied to its own evaluation method. Neither figure alone establishes how often the model will make errors in every real-world setting.

“deliberately crafted”

— Moonshot AI

Weights, Licence and Real-World Performance

The weights are not yet public, and the source material gives no published K3.1 licence. It is therefore unclear what users will be permitted to do with the weights if they arrive, including whether commercial use, modification or redistribution will be allowed. The Nov. 4 date is a company target, not confirmation that publication has occurred.

There are also limits to what can be concluded from the launch-day scores. Artificial Analysis evaluated defined benchmarks, while Moonshot’s demonstrations and several performance claims come from the company. The source material does not establish how the model performs across independent, real-world deployments, or whether its coding, CAD and finance examples can be reproduced consistently. The costs and hardware needed for self-hosting are also not specified beyond the model’s parameter count.

The Nov. 4 Weight Release Target

The next development to watch is whether Moonshot publishes K3.1 Pro’s weights by its stated Nov. 4 target, and what licence accompanies them. That information would clarify whether researchers and businesses can inspect or deploy the model outside Moonshot’s products, and under which conditions.

Further independent evaluations could test the company’s claims about game creation, CAD and finance, while broader use may show how benchmark results translate to practical work. Until then, K3.1 Pro is available through Moonshot’s apps, tools and API; its status as a downloadable open-weight model remains conditional.

Key Questions

Is Kimi K3.1 Pro available to download?

Not yet, according to the supplied report. Moonshot has made the model available through its apps and API and has promised to publish the weights by Nov. 4, 2026.

How did Kimi K3.1 Pro rank?

Artificial Analysis gave it an Intelligence Index score of 55.1, placing it fourth in the comparison provided. It scored below Opus 5.5, Sonnet 5.5 and Muse Wonder Full.

What does the model cost through the API?

Artificial Analysis listed prices of $2 per million input tokens and $10 per million output tokens. Actual spending depends on use, and those API prices do not represent local operating costs.

Is Moonshot’s 65% hallucination reduction independently confirmed?

No. That figure comes from Moonshot’s internal stress test, which the company said was designed to provoke errors. Artificial Analysis reported a smaller decline on a different test, from 53% to 44%; the results are not directly interchangeable.

What remains unknown about the open release?

The licence has not been published, and the weights have not yet been released. Their availability, permitted uses, release timing and practical self-hosting requirements remain unclear.

Source: ThorstenMeyerAI.com