Kimi K3.1 Pro: The AI Model Approaching The Frontier

🔍 Read the full analysis: Kimi K3.1 Pro: The AI Model Approaching The Frontier 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 in its apps, coding product and API on October 23, pitching it as a powerful model for coding, agentic work and knowledge tasks. In launch-day testing, Artificial Analysis gave it an Intelligence Index score of 55.1, fourth among the models listed, while the model’s weights are still promised—not available—for November 4.

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 puts it behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8 in the source’s October 24 comparison. The score reflects the evaluated model, not proof that the unreleased weights will reproduce the same results when run independently.

The service is priced at $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimated a cost of $2.11 per Intelligence Index task, compared with $5.98 for Opus 5.5 at its highest reasoning setting. These figures are evaluation-specific; they do not establish what every company or user would spend on a real workload.

Moonshot describes K3.1 Pro as a more heavily post-trained version of Kimi K3 and highlights demos involving game engines, 3D design and financial work. Those examples and several performance figures are Moonshot claims based on its demonstrations and internal tests. The separate Artificial Analysis results offer independent comparisons on selected benchmarks, but do not independently validate every vendor demo.

A Narrower Gap in Model Scores

The score places K3.1 Pro close to leading proprietary models on Artificial Analysis’s index: it is three points below the listed leader, rather than the 12-point difference cited for the best open-weight model in the comparison before this launch. That makes the result relevant to organizations weighing model capability against cost and control. The caveat is that the most consequential part of the pitch—access to downloadable weights—has not yet been delivered.

Artificial Analysis recorded a 67% score on Terminal-Bench 4.0, compared with 13% for Kimi K3, the strongest result in the source’s listed comparison. On the knowledge-work benchmark GDPval-AA, K3.1 Pro scored 1,846 Elo, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam it reached 52%, while Opus 5.5 scored 61%. Taken together, these results point to strong performance in some tasks, not uniform leadership across all uses.

The reported cost per index task is also lower than the figure for Opus 5.5, but a benchmark-derived cost is not a complete business case. Actual expenses depend on task mix, volume, latency requirements and deployment choices. If the weights arrive under terms that permit meaningful self-hosting, organizations could gain another option for handling sensitive workloads; neither the release terms nor the infrastructure burden are settled yet.

From Kimi K3 to K3.1

The source’s comparison shows Kimi K3 at 43.6 on the Intelligence Index and K3.1 Pro at 55.1. Other listed open-weight models, including Xiaomi’s MiMo-V2.6-Pro at 46.3 and DeepSeek V4.1 Flash at 45.1, scored below K3.1 Pro. However, the comparison labels K3.1 Pro’s weights as “promised by Nov 4,” while those other models are identified as open. Calling K3.1 Pro an open-weight leader is therefore conditional on a future release.

Moonshot’s internal hallucination test is another example of why benchmark claims need their source attached. The company reported a claim-level hallucination rate of 8.3%, down from 23.4% for Kimi K3, describing the reduction as about 65%. Artificial Analysis measured a different result on its AA-Omniscience test: 44% for K3.1 Pro, compared with 53% for Kimi K3. Moonshot said its own test was deliberately designed to provoke errors and did not represent overall error rates in daily use.

The model launched through Kimi apps, Kimi Work, Kimi Code and the Kimi API. Those hosted services are available now, but they are distinct from downloading and operating the model’s weights on a company’s own infrastructure.

“The hallucination 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 promised November 4 weight release has not happened yet, and the source reports that Moonshot had not published a licence for K3.1 Pro. Without the weights and terms, it is not possible to assess whether or how organizations may modify, redistribute or deploy the model themselves. The total parameter count also makes self-hosting a substantial infrastructure undertaking, rather than an automatic consequence of a weight release.

There are limits to what the reported results establish. Artificial Analysis evaluated the hosted model using its own benchmark suite; those scores do not confirm performance across every production task. Moonshot’s game, CAD and finance demonstrations remain vendor-provided examples, and its internal hallucination test is not a general measure of everyday error rates. The source material also does not specify how the model performs across a broad range of languages, workloads or deployment conditions.

The November Weight Release

The next milestone is Moonshot’s promised publication of K3.1 Pro’s weights by November 4. At that point, users and independent evaluators can check the files, licence and practical requirements for running the model, and test whether the benchmark results hold outside Moonshot’s hosted service. Until then, K3.1 Pro is available through Moonshot’s products and API, while its status as an accessible open-weight model remains pending.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a model from Moonshot AI, released on October 23, 2026, through the Kimi apps, Kimi Work, Kimi Code and the Kimi API. 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 gave it an Intelligence Index score of 55.1, fourth in the comparison reported on October 24. Its result was close to the listed leading scores, but it did not top every benchmark: for example, it scored 52% on Humanity’s Last Exam, compared with 61% for Claude Opus 5.5.

Can users download the model weights now?

No—not according to the source material. Moonshot said the weights would be published by November 4, and the report said no K3.1 licence had been published at the time. Both the release and its terms remain pending.

Are the 65% fewer hallucinations claim independently confirmed?

No. The roughly 65% reduction comes from Moonshot’s internal test, which compared claim-level hallucination rates of 23.4% for Kimi K3 and 8.3% for K3.1 Pro. Artificial Analysis reported a smaller change on its separate AA-Omniscience test, from 53% to 44%.

Source: ThorstenMeyerAI.com