🔍 Read the full analysis: Kimi K3.1 Pro: The Open-Weight Model Nearing The Frontier on ThorstenMeyerAI.com
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
TL;DR
Moonshot AI released Kimi K3.1 Pro in its apps and API on October 23, and Artificial Analysis scored it 55.1, placing it fourth among models in its Intelligence Index. The weights are promised by November 4, but have not been released, and Moonshot has not published the model’s licence.
Moonshot AI launched Kimi K3.1 Pro on October 23, with the model available through Kimi apps, Kimi Work, Kimi Code and the Kimi API. Artificial Analysis gave it a 55.1 Intelligence Index score, fourth among the models listed in its October 24 comparison; however, the weights are still pending and the licence has not been published, so it is not yet an available open-weight model.
Artificial Analysis ranked K3.1 Pro behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The evaluator’s table lists the Kimi weights as promised by November 4. Artificial Analysis said the model would lead open-weight models if its weights are published as promised; that ranking depends on a future release, not a download currently available to users.
The model is listed as a mixture of experts with 2.8 trillion total and 104 billion active parameters, a one-million-token context window and text and image input. Artificial Analysis reports API prices of $2 per million input tokens and $10 per million output tokens, with an estimated $2.11 per Intelligence Index task. Its comparison puts Claude Opus 5.5 at $5.98 per task, though cost estimates depend on the evaluator’s tasks and settings.
On Artificial Analysis’s Terminal-Bench 4.0 evaluation, K3.1 Pro scored 67%, compared with 13% for Kimi K3. The index also recorded 1,846 Elo on GDPval-AA, a simulated knowledge-work evaluation. Its Humanity’s Last Exam score was 52%, while the cited Opus 5.5 result was 61%. These results point to stronger performance on some tasks, not a general finding that it outperforms leading proprietary systems.
Infographic 01 · Claimed vs. measured
Moonshot says “matches the strongest.” Independent tests say: close.
The KPIs at the top are independent measurements. Vendor numbers and independent numbers come from different harnesses, so they are shown side by side, never mixed.
Moonshot reports
vendor claims own harness, partly internal tests
| Benchmark | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | vs. Claude Opus 5.5: 66.4% |
| BenchCAD with tools | 92.7% | vs. Claude Opus 5.5: 96.2% |
| OfficeQA Pro v2 | 64.4 | vs. Claude Opus 5.5: 62.2 |
| Kimi Hallu Bench Hard internal | 8.3% | vs. Kimi K3: 23.4% (about −65%) |
Artificial Analysis measured
independent highest reasoning setting per model
| Measurement | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 67% | vs. Claude Opus 5.5 (max): 60% |
| GDPval-AA | 1,846 Elo | vs. Claude Opus 5.5: 1,866 |
| Humanity’s Last Exam | 52% | vs. Claude Opus 5.5: 61% |
| AA-Omniscience hallucination rate | 44% | vs. Kimi K3: 53% · Gemini 4 Argon: 15% |
The 65% hallucination drop holds on Moonshot’s own stress test. On an independent knowledge test, the rate fell from 53% to 44%.Moonshot: the stress test rates “do not represent overall error rates in daily use.”
Sources: Moonshot AI, Kimi K3.1 Pro tech blog (23 Oct 2026) (left card, vendor charts); Artificial Analysis launch article (23 Oct 2026) and Artificial Analysis model page (data as of 24 Oct 2026) (KPIs and right card).
Powered by Thorsten Meyer AI
A Narrower Gap on Agent Tasks
The result matters because open-weight models offer the prospect of running model weights outside a vendor’s hosted service, giving organisations more control over deployment and data handling. K3.1 Pro’s benchmark score is close to the leading models in Artificial Analysis’s comparison, and its reported improvement on Terminal-Bench suggests a marked change from its predecessor on that evaluation.
But the practical significance is conditional. Until Moonshot releases the weights and states the licence, organisations cannot verify whether they can download, modify or deploy the model under terms suited to their needs. The reported 2.8 trillion total parameters also make self-hosting a substantial infrastructure undertaking; the benchmark ranking alone does not establish that running it locally will be affordable or straightforward.
Cost comparisons require similar care. Artificial Analysis estimates lower cost per index task for K3.1 Pro than for Opus 5.5, based on its own evaluation and the listed API rates. That does not mean every workload will cost one-third as much. Input and output volume, reasoning settings, task mix and service conditions can change the bill.
GPU workstation for running large language models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
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 the Artificial Analysis table, K3 scored 43.6 on the Intelligence Index, while K3.1 Pro scored 55.1. The comparison is a useful indication of measured change, but it does not by itself identify which training or system changes produced the difference.
Artificial Analysis’s October 24 figures compare models at their highest listed reasoning settings, where available. Its results are independent evaluations, while a number of Moonshot’s launch examples and claims come from the company’s own demonstrations or internal tests. The distinction matters: the two kinds of evidence measure different things and should not be treated as interchangeable.
For instance, Moonshot reported 8.3% claim-level hallucination on its internal Kimi Hallu Bench Hard, down from 23.4% for K3. The company described this as a roughly 65% reduction and said the test was deliberately designed to provoke errors, not to represent everyday error rates. Artificial Analysis measured a change from 53% to 44% on its separate AA-Omniscience evaluation. Those results use different tests and do not establish a single universal hallucination rate.
“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
NVIDIA RTX GPU for local AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Weights, Licence and Real-World Results
The largest open question is whether Moonshot will publish the weights by November 4, and what licence will govern their use. The source material does not provide a published K3.1 Pro licence. Until those details appear, developers and businesses cannot assess the legal permissions or restrictions attached to downloading and deploying the model.
Moonshot also promoted demonstrations involving game development, Blender and CAD, and finance. The supplied account describes these as company demonstrations or internal evaluations, and says the examples could not be independently verified. For example, Moonshot reported a 92.7% result on BenchCAD with tools, below the cited results for Claude Opus 5.5 and GPT-6 Astra. The exact conditions and broader reproducibility of the company’s demos remain unclear.
Benchmark scores are snapshots, not guarantees of performance on a company’s own data or workflow. Artificial Analysis’s evaluations provide external evidence for selected tests, but the available information does not establish how the model performs across all coding, research, financial or agentic tasks, or how its quality and costs compare in routine deployment.
As an affiliate, we earn on qualifying purchases.
The November Weight Release
The next stated milestone is Moonshot’s promised release of K3.1 Pro’s weights by November 4. Readers and prospective users will need to check whether the company meets that date, publishes a licence, and provides the files and documentation needed to run the model. The supplied material does not confirm a release plan beyond that promise.
After release, independent researchers and developers can test the weights directly, examine hardware requirements and compare results with the API version. That could clarify whether the benchmark performance is reproducible outside Moonshot’s hosted service and whether local deployment is viable for particular organisations. Until then, K3.1 Pro is a newly launched hosted model with promising independent scores, not a model already available for open deployment.
USB-C external SSD for AI model files
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is Kimi K3.1 Pro open-weight now?
No. Moonshot has promised the weights by November 4, 2026, but the source material says they had not been released as of October 24. The licence was also unpublished.
How did K3.1 Pro rank?
Artificial Analysis scored it 55.1 on its Intelligence Index, placing it fourth in the comparison reported on October 24, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full.
What does the benchmark say about coding and agent work?
Artificial Analysis recorded a 67% score on Terminal-Bench 4.0, compared with 13% for Kimi K3. That is a result on one benchmark and does not establish performance across all software or agent tasks.
How much does the API cost?
Artificial Analysis lists prices of $2 per million input tokens and $10 per million output tokens, and estimates $2.11 per Intelligence Index task. Actual costs depend on usage and task settings.
Can companies run it on their own hardware?
Not on the basis of the information available so far: the weights are pending, and no K3.1 Pro licence has been published. Artificial Analysis lists 2.8 trillion total parameters, indicating that self-hosting could require substantial infrastructure.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
