Kimi K3.1 Pro Nears The Frontier, Though Its Weights Aren’t Out Yet

🔍 Read the full analysis: Kimi K3.1 Pro Nears The Frontier, Though Its Weights Aren’t Out Yet on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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 October 23, presenting it as a leading open-weight model, but its weights have not yet been released. Artificial Analysis gave the model a 55.1 Intelligence Index score, fourth among the systems it evaluated, while Moonshot has promised the weights by November 4 without yet publishing a 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 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. The ranking reflects the models and evaluation settings listed by Artificial Analysis as of October 24, with the relevant models tested at their highest reasoning settings.

The evaluation gives K3.1 Pro a particular advantage on agentic tasks: it scored 67% on Terminal-Bench 4.0, compared with 13% for the earlier Kimi K3 in the same source’s table. Its score on Humanity’s Last Exam was 52%, compared with 47% for K3 and 61% for Opus 5.5. Artificial Analysis also recorded a lower hallucination rate on its AA-Omniscience evaluation: 44%, down from 53% for K3. These are results on named tests, not guarantees of performance in every use.

Artificial Analysis estimates K3.1 Pro costs about $2.11 per Intelligence Index task, compared with $5.98 for Opus 5.5 in its table. Listed API rates are $2 per million input tokens and $10 per million output tokens. Moonshot separately promotes coding, game-development, CAD and finance capabilities, including a claim of 65% fewer hallucinations on an internal test. That figure is not the result reported by Artificial Analysis, which measured a decline from 53% to 44% on its own test.

A Narrower Gap for Open Weights

The result could matter to businesses and developers weighing the performance of frontier models against the control that comes with running a model on their own infrastructure. If K3.1 Pro’s weights are released under terms that permit the needed uses, its score would put it close to the leading closed models in Artificial Analysis’s comparison. But the model is not yet downloadable, so that possibility remains conditional rather than a current option.

The reported cost per task is also relevant to teams comparing hosted API use. In Artificial Analysis’s test, K3.1 Pro’s estimated task cost was lower than Opus 5.5’s, though the figures apply to that benchmark setup and should not be treated as a universal price for a company’s workload. API rates, actual usage, reliability and the cost of operating a very large model locally are separate considerations.

For users, the distinction between independent benchmark results and vendor demonstrations is material. Moonshot’s examples may point to useful capabilities, but they do not establish that those results will reproduce across customers, tasks or production environments.

From Kimi K3 to K3.1 Pro

Kimi K3.1 Pro is described by Moonshot as a more heavily post-trained version of Kimi K3. In Artificial Analysis’s comparison, K3 scored 43.6 on the Intelligence Index, while K3.1 Pro scored 55.1. The increase is notable in the source’s agentic benchmark, where K3.1 Pro scored 67% on Terminal-Bench 4.0 and K3 scored 13%; the index and benchmark scores measure different tasks and should not be conflated.

Before K3.1 Pro’s launch, the source says the highest-ranked open-weights model in the Artificial Analysis comparison was Xiaomi’s MiMo-V2.6-Pro, at 46.3. K3.1 Pro’s higher score makes it a potential new leader in that category, but Artificial Analysis describes that status conditionally: it would lead once its weights are published. As of October 24, the model is available through Moonshot’s services, not as released weights for users to download.

Moonshot’s launch material also describes internal tests and demonstrations. It reports a 92.7% result on BenchCAD with tools, compared with 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra, and cites gains on its own finance evaluations. The company’s claimed 65% reduction in hallucinations comes from its Kimi Hallu Bench Hard test. Moonshot says that test is deliberately designed to provoke errors and that its rates do not represent overall error rates in ordinary use.

“The weights are promised “by Nov 4.””

— Moonshot AI

Weights, Licence and Real-World Results

The release date and usage terms remain unresolved. Moonshot has promised the weights by November 4, but as of October 24 they were not publicly available, and no K3.1 Pro licence had been published. Without the weights and licence, developers cannot confirm how the model can be deployed, modified or redistributed.

It is also unclear whether the weights, once released, will make local operation practical for typical organizations. The model has 2.8 trillion total parameters, and running it independently would require substantial computing resources. The source describes such deployment as a data-centre-scale undertaking; the actual hardware needs, operating costs and supported configurations have not been established here.

Benchmark results do not settle how the model will perform across real-world workloads. The independent evaluations provide comparable measurements on specific tests, while many of Moonshot’s examples and claims come from company demonstrations or internal benchmarks. Further independent testing, including in production settings, would help clarify their reach.

The November 4 Release Target

The next stated milestone is Moonshot’s promised publication of K3.1 Pro’s weights by November 4. Developers and organizations will then be able to check the release itself, read the licence and assess whether the terms and hardware demands fit their needs. The company’s target is not a guarantee that the release will happen on that date.

Further independent evaluations could test whether the benchmark advantage carries over to other coding, knowledge and agentic tasks, and whether Moonshot’s demos can be reproduced outside its own environment. Until the weights and licence are available, K3.1 Pro remains a model accessible through Moonshot’s services rather than a released open-weight model. AI model performance and service costs can vary by task and usage; benchmark scores and listed prices do not guarantee results or total costs.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a model from Moonshot AI, described as a more heavily post-trained version of Kimi K3. Moonshot made it available in its apps, Kimi Work, Kimi Code and API on October 23, 2026.

How did it rank on Artificial Analysis?

Artificial Analysis gave it a 55.1 Intelligence Index score and placed it fourth in its comparison as of October 24, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.

Are Kimi K3.1 Pro’s weights available to download?

No. Moonshot had not published the weights as of October 24 and said they were promised by November 4. The model was available through the company’s products and API.

Has Moonshot published a licence for the weights?

No K3.1 Pro licence had been published in the supplied information. The terms will matter for anyone considering downloading, modifying or redistributing the model after its weights are released.

Is the 65% hallucination reduction independently verified?

The 65% figure comes from a Moonshot internal test. Artificial Analysis reported a drop from 53% to 44% on its separate AA-Omniscience evaluation. Moonshot also cautioned that its internal test was designed to provoke errors and did not represent ordinary-use error rates.

Source: ThorstenMeyerAI.com