Kimi K3.1 Pro is Moonshot AI's newest model, announced on 24 October 2026 as "our most capable 3T-class open-weight model to date."[1] The word "open-weight" is doing future work: the same page says the full weights "will be released by Nov 4," and the K3.1 licence is not stated. This article separates what is documented from what is promised, and reads the licence that governs the previous model, Kimi K3, because it is the only licence text available to read today.
What is documented today
Moonshot's announcement lists the model for Kimi.com, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API. The API names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed, the latter described as delivering "2-3× the speed" of K3.1 Pro. The context window is one million tokens, and Zero Data Retention is available for API customers. Enterprise access is off by default at launch and must be enabled by admins.[1]
The announcement does not give a parameter count, a price or a licence. We tried Moonshot's API pricing documentation, but the price table did not load in the content we could read, so we state no price. The documentation does say that K3-series models share a 1,048,576-token context window and that cache writes are billed separately by time-to-live.
For scale, the published Kimi K3 model card lists 2.8 trillion total parameters with 104 billion activated per token, a mixture-of-experts design with 896 experts of which 16 are selected per token, and weights trained with MXFP4 quantisation.[2] Moonshot does not say K3.1 Pro's architecture beyond the "3T-class" description, so treat K3's figures as context, not as K3.1's specification.
Our arithmetic from K3's numbers: 104 billion of 2.8 trillion is about 3.7% of parameters active per token. Compute per token therefore scales with the active share, but memory does not: all experts must be held somewhere, so a model like this is a multi-GPU deployment regardless of how few parameters fire per token.
The benchmark table is the vendor's
Moonshot's announcement reports the following. We could read the figures from the page text; the table itself is an image, so some rows may have additional context we could not see.[1]
| Benchmark | Kimi K3.1 Pro | Comparison reported by Moonshot |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | GPT-6 Astra 58.2%; Claude Opus 5.5 66.4% |
| Kimi Code Bench 3.0 (long-horizon) | 61.2% | Kimi K3 55.95% |
| BenchCAD, with tools | 92.7 | Claude Opus 5.5 96.2; GPT-6 Astra 95.9 |
| BenchCAD, without tools | 72.8 | Claude Opus 5.5 73.0 |
| Kimi Hallu Bench Hard, with tools (claim-level hallucination rate, lower is better) | 8.3% | Kimi K3 23.4%; Claude Opus 5.5 15.0%; GPT-6 Astra 5.5% |
| xdev-bench-hard (pass@3) | 81.0% | Kimi K3 33.3% |
Default settings for most results were reasoning effort "max," temperature 1.0 and top-p 0.95.[1]
How to read it:
- Several rows are Moonshot's own benchmarks. Kimi Code Bench and Kimi Hallu Bench are named after the company, and the page also lists internal finance and business-operation evaluations. We have no independent description of their construction.
- The mixed result is the honest part. On Terminal-Bench 4.0 and BenchCAD the table shows Claude Opus 5.5 ahead or level, and on the hallucination benchmark GPT-6 Astra is lower. A vendor table that is not all wins is more informative than one that is.
- Run settings differ from yours. "Max" reasoning effort costs tokens and time. A result at max effort does not predict cost or latency at the setting you would deploy. Our guide to reading an AI benchmark claim lists the questions to ask of any such table.
A cyber-capability cap, disclosed
The most unusual passage is about security. Moonshot says it removed cyber-related data from the mid-training and reinforcement-learning stages for the open-weight release, and writes: "we aim not to release open weights that would extend the absolute frontier of cyber capability." It adds that K3.1 Pro's offensive cyber capability is "by design" below what its pipeline can produce, and that it remains behind frontier proprietary models that have been publicly available for at least three months.[1]
The page gives one number to support this: on xdev-bench-hard, a checkpoint without these measures scored 90.5%, against 81.0% for the final model.[1] That is a company-reported before-and-after on a single benchmark. It shows the intervention changed the score; it does not show how much real-world capability was removed, and fine-tuning open weights later can change the picture. For how labs frame such limits more generally, see our comparison of frontier safety frameworks.
The licence you can read today
Kimi K3's weights and code are released under the Kimi K3 License.[2] The text is short and has these elements (our paraphrase, with quoted phrases from the file; this is not legal advice):[3]
- Broad grant. Use, modification, fine-tuning and redistribution of the weights, code and documentation, with the copyright and permission notice included in copies, and compliance with applicable law.
- Model-as-a-service threshold. Giving a third party access to model inference or fine-tuning, "e.g., via API," when the licensee and its affiliates have aggregate revenue above 20 million US dollars requires a separate agreement with Moonshot AI.
- Large-product attribution. Products with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, must display "Kimi K3" prominently in the user interface.
- Exemptions. Internal use that does not make the model or its outputs available to third parties, and use through Moonshot's official products or certified inference partners.
Compare this with the distinction we draw in open weights versus open source: a licence can allow download and modification and still restrict commercial offerings. Whether K3.1 Pro ships under the same terms is not stated on the announcement page, so verify it when the weights are published.
What to do now
- Evaluating through the API: run your own tasks at the reasoning-effort setting you would deploy, and measure cost and latency, not only accuracy. Check Moonshot's pricing page directly; we did not verify a price.
- Planning to self-host: wait for the weights and the model card. A 3T-class model is a multi-node deployment; our inference serving explainer covers the memory side.
- Planning to resell inference: read the licence, and ask Moonshot about the revenue threshold before you design a product around the weights.
- Judging the claims: discount the Kimi-named benchmarks until someone independent reproduces them.
What remains uncertain
We have not tested K3.1 Pro. The announcement states no launch time beyond "rolls out today," so we use the 24 October date shown on Moonshot's blog listing. The weights, licence, parameter count and price for K3.1 are unconfirmed at the time of writing (retrieved 2026-10-27), and the Hugging Face organisation page we read showed Kimi-K3 as its most recently updated model.[4] We will add a dated changelog line when the weights and licence are published.
