What the new model offers
Beijing-based Moonshot AI unveiled Kimi K3.1 Pro on 23 October 2026. It is the first model in the company's new K3.1 series and, to date, its strongest open-weight model in the 3-trillion-parameter class. The model is a post-training-improved version of Kimi K3, released in July 2026, and its development used considerably more post-training compute than previous projects. Development covered hundreds of subfields.
The new model's context window is 1 million tokens. This means the model can work at once with very large volumes of text and data. Moonshot also released the Kimi K3.1 Pro Highspeed variant. According to the company, this variant runs two to three times faster than the standard model. The speed difference may shorten waiting time for users, but this figure, too, is based on the company's own statement.
According to Bloomberg's reading, this launch is Moonshot's first major update since Kimi K3. Kimi K3 took the Beijing startup into the ranks of leaders in artificial intelligence during the summer months. The new model is aimed at areas such as financial research, programming and 3D modelling. According to a Bloomberg source, the company is valued at around 50 billion dollars.
Why cyber capabilities were limited
Moonshot openly stated that it deliberately reduced the new model's cyberattack capabilities. To do this, the company removed cyber-related data from the model's training corpus and also removed it from reinforcement-learning environments. At the same time, the model's general programming skills were preserved. In other words, the restriction aims not to remove the model's ability to write code entirely, but specifically to narrow its capabilities directed at cyberattacks.
The company explained its position as follows:
We aimed not to release open weights that expand the absolute frontier of cyber capabilities.
According to Moonshot's internal cyber benchmark, K3.1 Pro scored 81.0 percent. A comparable variant without risk-reduction measures scored 90.5 percent. This difference shows that the restriction measures lowered the model's result in the cyber field. At the same time, these results are given according to company data and have not been independently verified.
What the benchmark results show
According to the company, independently unverified benchmark results show that Kimi K3.1 Pro is strong in programming. In the Terminal-Bench 4.0 programming-agent test, the model scored 65.7 percent. In the same test, OpenAI GPT-6 Astra scored 58.2 percent and Anthropic Claude Opus 5.5 scored 66.4 percent. This comparison puts Kimi K3.1 Pro ahead of Astra and close to Opus 5.5. However, these figures are also based on the company's own reports and have not been independently verified.
In the BenchCAD 3D-model programming test, the situation is different. According to the company, based on independently unverified results, Kimi K3.1 Pro scored 92.7 percent. In the same test, Claude Opus 5.5 scored 96.2 percent and GPT-6 Astra scored 95.9 percent. So in 3D modelling the new model lags behind its rivals. Together, the two tests distinguish the model's strengths and weaknesses: while it matches the leaders on programming-agent tasks, it still lags in 3D-model code.
Caution is needed when reading benchmark figures. They come from the company's own reports and have not been independently verified. They should therefore be taken not as a final conclusion, but as initial indicators presented by the company. Once independent evaluations are published, these comparisons may change.
When and where it can be used
Both Kimi K3.1 Pro and K3.1 Pro Highspeed are already available via Kimi.ai, Kimi Work, Kimi Code and Kimi API. This means users can access the model via the web, a work environment, a programming tool and an application interface. Moonshot plans to publish the model's full open weights by 4 November 2026. Once the open weights are published, other developers will be able to study the model and use it in their own systems.
The model's open-weight nature makes its wide distribution easier. Moonshot therefore stresses that it took the decision to limit cyber capabilities before the open weights are published. In the company's view, an openly distributed model should not expand the frontier of cyberattacks. This approach looks like an attempt to find a balance between openness and safety.
Against the regulatory backdrop in China
The model's launch coincided with a period when new regulatory initiatives in artificial intelligence are being discussed in China. Last week, China proposed a draft export-control measure that would ban the export of technologies for creating malware with artificial intelligence and restrict tools for identifying vulnerabilities with artificial intelligence. The reason given for the draft was risks to critical infrastructure and state systems.
This background helps to understand the context of Moonshot's decision. On the one hand, the company is releasing a powerful open model. On the other hand, it is announcing that it is limiting that model's cyberattack capabilities. The export restrictions proposed by the Chinese government are also forming precisely around such technologies. Thus, the launch of Kimi K3.1 Pro has become part of a broader discussion not only about a technical update, but also about the safety of artificial intelligence models and their distribution.
