🔍 Read the full analysis: Kimi K3.1 Pro: A Frontier-Level Showing Without Released Weights on ThorstenMeyerAI.com
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
TL;DR
Moonshot AI released Kimi K3.1 Pro through its apps and API on October 23, describing it as a 3-trillion-parameter open-weight model. Artificial Analysis scored it 55.1 on its Intelligence Index, while the model’s weights are promised by November 4 and no licence has been published.
Moonshot AI launched Kimi K3.1 Pro on October 23, and an evaluation by Artificial Analysis places it fourth on its Intelligence Index, with a score of 55.1. The company says the model’s weights will be released by November 4, but they are not yet public and a licence for K3.1 has not been published, so users cannot yet verify its performance by running those weights themselves.
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. It became available through Kimi apps, Kimi Work, Kimi Code and the Kimi API. The service lists prices of $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimates a cost of about $2.11 per Intelligence Index task.
On the index, K3.1 Pro scored 55.1, behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0, and Muse Wonder Full at 55.8. Artificial Analysis says the model would lead its open-weight ranking once the weights are published. That is a conditional assessment: at present, the index classifies the weights as promised, not available. On Terminal-Bench 4.0, its evaluation recorded a 67% score, compared with 13% for the earlier Kimi K3.
Moonshot’s launch materials also promote work involving game engines, Blender, computer-aided design and finance. Those examples and several headline performance claims come from the company’s own demonstrations or internal tests, rather than independent verification. On Artificial Analysis’s AA-Omniscience evaluation, the measured hallucination rate fell from 53% for K3 to 44% for K3.1 Pro; that result differs from Moonshot’s internal claim of a roughly 65% reduction on a separate test.
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 to Closed Models
The result puts Kimi K3.1 Pro close to several leading proprietary models on one independent benchmark, while its listed API price is lower than that of Claude Opus 5.5. Artificial Analysis estimates about $2.11 per index task for Kimi against $5.98 for Opus 5.5. Those figures offer a benchmark-based comparison, not a guarantee that every real-world workload will have the same cost or quality.
The potential change for companies depends on the promised weights becoming available under a usable licence. If that happens, businesses and researchers could examine and deploy the model in ways that an API-only service may not permit. But the scale matters: a model listed at 2.8 trillion total parameters is not a routine download for most organisations, and self-hosting it could require substantial data-centre resources. Until publication, claims about independent deployment remain prospective.
For users choosing a hosted model, the launch adds another benchmark contender and a new API option. It does not settle which model is best for a particular job. The index combines selected evaluations, and the gap between benchmark results and everyday performance can matter for coding, research or office tasks. Moonshot’s demonstrations may point to intended uses, but they should not be treated as proof that the model will reliably complete those tasks in production.
Products Worth Considering
Apple MacBook Pro 14.2-inch with M3 Max, 14-Core CPU, 30-Core GPU, 36GB RAM, 1TB SSD, Space Black
Apple MacBook Pro 16.2-inch (M5 Pro, 48GB RAM, 2TB SSD, Space Black)
Apple MacBook Pro 14‑inch (2021) with M1 Pro, 16 GB RAM, 512 GB SSD – Space Gray
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. The benchmark table supplied by Artificial Analysis shows K3 at 43.6 on the Intelligence Index and K3.1 Pro at 55.1. That is a result on the organisation’s index, not a general measure of improvement across all tasks.
Before this launch, Artificial Analysis listed Xiaomi’s MiMo-V2.6-Pro at 46.3 as the highest-scoring open model in the comparison. K3.1 Pro’s score is higher, but the evaluator’s open-weight framing depends on the promised release. Its table also shows differing strengths across models: K3.1 Pro scored 67% on Terminal-Bench 4.0, while knowledge-evaluation results were less dominant. On Humanity’s Last Exam, K3.1 Pro scored 52%, compared with 61% for Opus 5.5.
Moonshot’s own account describes a long-running internal simulation in which the model worked on a walking simulator, along with examples of game development and parametric modelling. The company also reports a 92.7% result on BenchCAD with tools, below the cited scores for Claude Opus 5.5 and GPT-6 Astra. These company-reported demonstrations provide context for Moonshot’s product pitch, but they are separate from the independent Artificial Analysis index.
“The weights are promised “by Nov 4.””
— Moonshot AI
Weights, Licence and Real-World Accuracy
The most immediate open question is whether Moonshot will publish the weights by November 4, and under what terms. No K3.1 licence had been published in the source material. Without the weights and licence, users cannot establish how freely the model can be modified, redistributed or used in commercial products.
It is also unclear how much of the launch’s showcased work will generalise beyond demonstrations and controlled evaluations. Moonshot’s claim of a roughly 65% reduction in hallucinations comes from its internal Kimi Hallu Bench Hard test, comparing 23.4% for K3 with 8.3% for K3.1 Pro. Artificial Analysis recorded a smaller change on a different evaluation, from 53% to 44%. The tests have different methods, so their results are not directly interchangeable; neither figure alone establishes the model’s error rate across everyday use.
Finally, benchmark scores do not answer the operational questions facing organisations that might self-host the model. The available information does not specify the hardware configuration needed to run the full system, its deployment cost, or how performance would change under different workloads. Those details will matter alongside benchmark rankings and API prices.
The November 4 Release Test
The next stated milestone is Moonshot’s promised November 4 weights release. Publication would allow outside users to examine the files and assess the accompanying licence, though availability alone would not confirm every vendor claim or make the model practical to run on ordinary hardware.
Readers should also watch for independent evaluations after release, including whether third parties reproduce the launch-day benchmark results and test the model on their own tasks. For now, the confirmed picture is a strong score from Artificial Analysis and live access through Moonshot’s products and API; the open-weight status, licence terms and broader real-world performance remain to be established.
Key Questions
Is Kimi K3.1 Pro open-weight now?
No. Moonshot released access through its apps and API on October 23, but the model weights have not yet been published. The company said they are due by November 4.
How did Kimi K3.1 Pro rank?
Artificial Analysis gave it a score of 55.1 on its Intelligence Index, placing it fourth among the models listed in the October 24 comparison. The evaluator says it would lead the open-weight ranking if its weights are released.
What licence will the weights use?
The source material says no licence for K3.1 had been published. The terms are therefore still unknown, including what uses or redistribution may be allowed.
Is Moonshot’s 65% hallucination reduction independently verified?
No. That figure comes from Moonshot’s internal test, which the company says was designed to provoke errors and does not represent everyday error rates. Artificial Analysis measured a reduction from 53% to 44% on a separate evaluation.
Can organisations run Kimi K3.1 Pro on their own hardware?
That cannot be determined from the launch information alone. Artificial Analysis lists 2.8 trillion total parameters, and the source describes self-hosting at that scale as a data-centre project. Specific hardware and deployment requirements were not provided.
Source: ThorstenMeyerAI.com
