Kimi K3.1 Pro: What Moonshot Says About Its 3T Open-Weight Model, and What to Test Before Switching

Moonshot AI released Kimi K3.1 Pro on October 23. It calls the model its most capable 3T-class open-weight model, with a 1-million-token context window. The weights are not out yet: Moonshot says the full weights will be released by November 4. Every benchmark number below comes from Moonshot's own launch post, and I have not run the model.

I read one primary source: Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence. I read the main sections and part of the footnotes. I did not read the later benchmark tables in full, the API docs or any independent evaluation.

What was released

Per the post, K3.1 Pro is the first model in the K3.1 series and a post-trained version of Kimi K3, with substantially larger post-training compute. A second model, Kimi K3.1 Pro Highspeed, is said to run 2 to 3 times faster than K3.1 Pro.

Both are available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API under the names kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot says Zero Data Retention is available for API customers and that Kimi Enterprise access is off by default.

The benchmark claims

Moonshot says that across coding, agentic and knowledge benchmarks, K3.1 Pro "matches or approaches" the strongest proprietary models. Examples from the post:

Benchmark	K3.1 Pro	Comparison given 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 (internal)	61.2%	Kimi K3 55.95%
BenchCAD, with tools	92.7%	Claude Opus 5.5 96.2%, GPT-6 Astra 95.9%
OfficeQA Pro v2	64.4	Described as state of the art

Note what the footnotes say. K3.1 Pro was run with the Claude-Code harness, at max reasoning effort. Other vendors' scores are mostly taken from their own posts or public leaderboards, and Moonshot adds that proprietary model testing "may involve some fallback behavior". The Kimi Code Bench is Moonshot's own. These are not controlled head-to-head runs.

Factual accuracy

Moonshot reports that on its internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. The post itself says the examples are deliberately crafted to probe weak spots and that the rates do not represent overall error rates in daily use. Take that caveat at face value.

The cyber-capability approach

This is the most unusual part of the post. Moonshot says it found cyber-related content in its general mid-training corpus and cyber tasks in its reinforcement learning environment mix, and that removing them at both stages reduced vulnerability-research capability. It reports on its in-house xdev-bench-hard:

• A checkpoint without the removal measures solved 90.5% of problems at pass@3.
• The final K3.1 Pro reaches 81.0%.

Moonshot states a principle of not release open weights that extend the absolute frontier of cyber capability, and that K3.1 Pro remains behind frontier proprietary models that have been publicly reported. It also says it intends to progressively raise the cyber capability available in open weights. The benchmark is in-house and I could not check it. A drop from 90.5% to 81.0% is a reduction, not removal, and the post says as much.

Availability and what is missing

• Weights: promised by November 4, 2026. Until they are published, "open-weight" describes an intent.
• Licence: I did not find licence terms in the parts I read.
• Pricing: I did not find pricing in the parts I read.
• Hardware needs for self-hosting a 3T-class model: not covered in what I read.

What I would test

These are my own untested suggestions.

1. Run your own tasks. Use your real repos and tickets, not public benchmarks, and compare against the model you use now.

2. Check the harness. The scores use the Claude-Code harness. See how it behaves in the tool you actually use.

3. Measure long context. A 1M-token window is a claim about capacity, not about quality at the end of the window.

4. Compare Highspeed and Pro on both latency and accuracy for your workload.

5. Wait for the weights and the licence before planning self-hosting, and check that the released files match the announcement.

6. Check data terms for the API if you send private code, including Zero Data Retention eligibility.

What I would not conclude

• That it beats Claude Opus 5.5 or GPT-6 Astra. On the post's own Terminal-Bench figure it is slightly behind Opus 5.5.
• That the benchmark gains come only from the model. Harness and effort settings matter.
• That the open-weight release is guaranteed on the stated date.
• That cyber risk is solved. The post describes a reduction, measured on an internal test.

Frequently asked questions

What is Kimi K3.1 Pro?

Moonshot AI's new flagship model in the K3.1 series, a post-trained version of Kimi K3 with a 1-million-token context window, per its launch post.

Are the weights available?

Not yet. Moonshot says they will be released by November 4, 2026.

What is Kimi K3.1 Pro Highspeed?

A variant Moonshot says is 2 to 3 times faster than K3.1 Pro.

Where can I use it today?

Kimi.ai, the Kimi apps, Kimi Work, Kimi Code and the Kimi API, per the post.

How does it compare with Claude Opus 5.5?

On Terminal-Bench 4.0 Moonshot reports 65.7% against 66.4% for Opus 5.5. The comparison is Moonshot's, with the setup described in its footnotes.

What did Moonshot do about cyber capability?

It says it removed cyber-related data from mid-training and filtered cyber RL environments, and reports a drop on its internal benchmark from 90.5% to 81.0%.

Should I switch my coding agent to it?

I cannot say from a launch post. Test it on your own tasks first.

Sources

• Moonshot AI, Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence (October 23, 2026): release details, benchmark claims, cyber approach, availability and footnotes. Later benchmark tables were not read in full.