Beijing-based Moonshot AI introduced Kimi K3.1 Pro on Saturday, October 24, calling it the company’s most capable 3-trillion-parameter-class open-weight model to date. The company plans to publish the model’s full weights by November 4, so outside developers can download, run, and modify the files on their own systems. Before that release, Moonshot said, it deliberately reduced the model’s ability to conduct offensive cyber operations.
An open-weight model differs from a service like ChatGPT in one practical way: the weights are the model itself, distributed as files, and once they spread, the original developer loses much of its control over who uses them and what safeguards survive. That is why Moonshot’s announcement drew attention. The company removed cyber-related data from the model’s training corpus and from its reinforcement-learning environments, while keeping its general coding skills intact.
“We aim not to release open weights that would extend the absolute frontier of cyber capability,” the company said, according to Reuters.
Moonshot offered numbers to back the claim. On its internal xdev-bench-hard test, which asks a model to find and exploit software vulnerabilities in a controlled setting, the release version scored 81.0%. An unfiltered version of the same training pipeline scored 90.5%. The predecessor model, Kimi K3, scored 33.3% on the same test. The finished model is therefore weaker than what Moonshot’s pipeline could produce, and far stronger than the model it replaces.
Those numbers come with a limit the company does not hide: they have not been independently verified. The benchmark, the test selection, and the comparisons were assembled by Moonshot, so the percentages read as company evidence until outside researchers run their own tests. November 4 is when that becomes possible.
A model built for code, with a long memory
Outside the security question, K3.1 Pro is a large general-purpose model. It carries a 1-million-token context window, enough to take in very long documents or codebases in a single pass, and Moonshot sells a Highspeed variant it says runs two to three times faster than the standard version. Both are available now through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, according to the company’s announcement.
On company-reported benchmarks, K3.1 Pro scored 65.7% on Terminal-Bench 4.0, a test of coding agents on real terminal tasks, against 58.2% for GPT-6 Astra and close to Claude Opus 5.5’s 66.4%. Independent tracker Artificial Analysis gave the model a score of 55, ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon at 53 each, and behind Claude Opus 5.5 at 58. Simon Willison’s report put its price at $2 per million input tokens and $10 per million output tokens, matching GPT-6.1 Sol and half the price of Claude Opus 5.5.
Moonshot also reported a drop in its own factual-accuracy stress test, with the claim-level hallucination rate falling from 23.4% for K3 to 8.3%. The company cautioned that the test is a deliberate stress test and does not represent error rates in daily use.
Why governments are watching the same question
Moonshot’s announcement came shortly after China proposed export controls that would ban exports of AI-driven malware-generation and mutation technology and restrict AI-assisted vulnerability-discovery tools. The draft measures remain open for public comment, and they cite the risk that such capabilities could be used against critical infrastructure, defense-industrial systems, and core government systems.
The concern is not abstract. Models that can write code, operate software tools, and complete long, multi-step tasks can help security teams find weaknesses before criminals do. The same abilities can automate parts of an attack. Moonshot’s own argument for releasing the weights anyway is that open models let more organizations inspect and secure their own software, including older systems that receive little attention.
Chinese AI models have already drawn scrutiny in Washington. The United States ordered DeepSeek out of federal civilian agencies earlier this month, including tools built on its models, as Chip Current reported. K3.1 Pro will face a different kind of scrutiny: a public test by researchers with the weights in hand.
That test arrives November 4. If the safeguards hold up outside Moonshot’s lab, the company will have shown that a developer can ship a near-frontier model and leave some of its sharpest edges off. If they do not, the release will show how hard that restraint is to keep once a model belongs to everyone.
