Direct answer: Moonshot AI introduced Kimi K3.1 Pro on October 23, 2026, as the first model in its new K3.1 series. It is a large mixture-of-experts model in the 3-trillion-parameter class, with a 1-million-token context window, available now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Its full model weights are promised by November 4, 2026, but they have not been published yet, so for now it is an open-weight model in waiting rather than one you can download and run yourself. The most unusual part of the launch is deliberate: Moonshot reduced the model’s cyber-offensive ability on purpose, removing cyber-related data from its training, while keeping its general coding skills intact.
What Is Kimi K3.1 Pro?
Kimi K3.1 Pro is Moonshot AI’s most capable model to date, and an upgraded version of the earlier Kimi K3. Moonshot, the Beijing-based startup behind the Kimi chatbot, describes it as its most capable 3T-class open-weight model, built with significantly more computing power during post-training than its predecessor. Independent evaluator Artificial Analysis lists it as a mixture-of-experts model with 2.8 trillion total parameters, of which 104 billion are active at any one time, and it accepts both text and image input.
That distinction matters. Total parameters are not the same as active parameters. A mixture-of-experts design routes each task to a subset of specialist sub-models, which is how Kimi K3.1 Pro can be very large on paper while still being practical to serve through an API. For users, the more relevant numbers are the ones that affect real work: a 1-million-token context window, strong coding results, and a faster Highspeed variant for people who need quicker responses.
Key Facts at a Glance
- Launched: October 23, 2026, across Kimi.ai, Kimi Work, Kimi Code and the Kimi API
- Model type: Mixture-of-experts, 3-trillion-parameter class (2.8 trillion total, 104 billion active, per Artificial Analysis)
- Context window: 1 million tokens
- Input: Text and images
- Variants: Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed, which Moonshot says runs two to three times faster with the same capabilities
- Open weights: Promised by November 4, 2026; licence not yet published at launch
- API pricing listed by Artificial Analysis: $2 per million input tokens and $10 per million output tokens
What Kimi K3.1 Pro Is Built to Do
Moonshot pitches Kimi K3.1 Pro as a general-purpose agent model rather than a narrow chatbot. According to the company, it delivers substantial improvements across coding, 3D creation, financial research and other agentic capabilities, and it covers the full range of coding work, from quick fixes in real codebases to full-project builds. It is also positioned for end-to-end financial workflows, long-running autonomous tasks and knowledge-based work.
That agentic focus is where the model looks strongest in independent testing. On Terminal-Bench 4.0, a benchmark that tests coding agents on real terminal-based tasks, Moonshot reports a score of 65.7 percent, ahead of GPT-6 Astra at 58.2 percent and close to Claude Opus 5.5 at 66.4 percent. Artificial Analysis measured a similar result, 67 percent, against just 13 percent for the earlier Kimi K3, which suggests the jump is real even if exact scores vary with test settings. That pattern is familiar in AI benchmarking: different setups produce slightly different numbers, but a large gap in the same direction is worth paying attention to.
On broader capability, Artificial Analysis gave Kimi K3.1 Pro an Intelligence Index score of 55.1 in its October 24 comparison, placing it fourth, behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. In other words, it sits three points behind the leading proprietary model, and Artificial Analysis says it would lead open-weight models once its weights are published. That last point is conditional, not current: until November 4, nobody outside Moonshot can run the weights independently.
The Cyber Capability Decision: What Moonshot Actually Changed
The headline that sets this launch apart is not a benchmark. It is a safety decision Moonshot made before release.
According to Reuters, Moonshot removed cyber-related data from the training corpus and reinforcement-learning environments used to develop Kimi K3.1 Pro, while retaining general coding skills. The company said: We aim not to release open weights that would extend the absolute frontier of cyber capability. On its own internal cyber benchmark, which involved finding and exploiting software vulnerabilities in a controlled environment, Kimi K3.1 Pro scored 81.0 percent, compared with 90.5 percent for a similar version without the risk-reduction measures. Moonshot said the comparison showed that filtering cyber-related data reduced offensive cyber performance, although the model still outperformed its predecessor. Those internal results have not been independently verified.
There is an important nuance here. Reducing cyber training data does not make a powerful coding model harmless, and Moonshot does not claim otherwise. Artificial Analysis still measured a 64 percent score on its Cyber Index, the highest publicly accessible result in that comparison and second overall behind a restricted-access model at 69 percent. General coding ability, long context and agentic tool use can all be applied to security work, legitimate or otherwise. What Moonshot has done is lower the ceiling on purpose, not remove the capability entirely.
The context helps explain why. Governments, technology companies and security researchers are increasingly concerned that AI models which can write code, operate software tools and complete long, multi-step tasks could make sophisticated hacking easier and cheaper. China proposed export controls last week that would ban exports of AI-driven malware-generation technology and restrict AI-assisted vulnerability-discovery tools, with the draft measures open for public comment. At the same time, Moonshot argues open-weight models can improve security by helping more organisations inspect and secure their own software, including older systems that receive little attention. Both things can be true: wider access helps defenders, and it also makes central oversight harder once weights are widely distributed.
That tension is also visible elsewhere in the industry. Anthropic recently launched a free security scanning programme for open source projects, while NVIDIA has been building safety architecture for AI agents, and both efforts reflect the same problem: as agents get more capable, the controls around them have to keep up. You can read more about those parallel efforts in our coverage of Anthropic Cyber Mission and NVIDIA OpenShell.
Kimi K3.1 Pro vs the Competition
It helps to place Kimi K3.1 Pro against the models people are actually choosing between right now. On coding-agent tasks, it is close to the best proprietary models. On broad intelligence scores, it trails them by a small margin. On cost per task, Artificial Analysis estimated $2.11 per Intelligence Index task for Kimi K3.1 Pro, compared with $5.98 for Claude Opus 5.5 at its highest reasoning setting, though those figures reflect that benchmark’s pricing and task assumptions, not a universal cost for every workload.
The comparison with other open-weight models is where the launch matters most. Mistral Large 4, for example, was also launched as a very large open-weight model with its weights following later, and the pattern is becoming familiar: announce first, publish weights weeks later, and let developers judge the model in the hosted product while they wait. Our explainer on Mistral Large 4 covers how that staggered release worked in practice. For Kimi K3.1 Pro, the November 4 date is the moment to watch, along with the licence terms, which had not been published at launch.
For everyday users, the more immediate comparison is with proactive agent products such as Claude Duet, Anthropic’s agent that works in the background. Kimi K3.1 Pro is a model, not a finished personal assistant product, so the right question is not which chatbot feels nicer today, but which model developers and businesses can build on, and under what licence, once the weights arrive.
How to Try Kimi K3.1 Pro Now
If you want to test it before the weights release, the hosted route is the only option right now:
- Go to Kimi.ai and select Kimi K3.1 Pro, or Kimi K3.1 Pro Highspeed if you need faster responses and are willing to pay more for speed.
- Try it on a real task from your own work, not a demo prompt: a multi-file coding fix, a long document you need summarised with sources checked, or a financial research task with numbers you can verify.
- Check its work carefully. Long context is useful, but it does not guarantee accuracy, and agentic models can still make confident mistakes across many steps.
- If you plan to self-host, wait for the November 4 weights release and read the licence before budgeting hardware or building a product around it.
For developers, the Kimi API is the other route in. Artificial Analysis lists API pricing at $2 per million input tokens and $10 per million output tokens, which makes it cheaper per token than the leading proprietary models in the same comparison, though your real cost will depend on how much reasoning and context each task uses.
What to Watch Before November 4
Three things will decide whether Kimi K3.1 Pro lives up to its launch claims. First, the weights release itself: does it arrive on time, in full, and with a licence that genuinely allows commercial use and modification? Second, independent testing once developers can run the model themselves, rather than through Moonshot’s hosted service. Third, whether the cyber-capability reduction holds up in practice, or whether fine-tuning and prompting recover much of the removed ability. Open-weight releases are hard to recall once distributed, which is exactly why Moonshot’s pre-release filtering is worth noting, and why it will also be tested by the community.
None of that changes the basic picture today: this is a strong, near-frontier open-weight model in hosted form, with a deliberately lowered cyber ceiling, a very large context window, and a weights release that is still a promise rather than a download. If you are choosing a model for coding or agentic work, it is worth testing now. If you are choosing a model to self-host, wait until November 4 and judge the actual release, not the announcement.
Frequently Asked Questions
When was Kimi K3.1 Pro released?
Moonshot AI introduced Kimi K3.1 Pro on October 23, 2026, across Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Reuters covered the launch on October 24, 2026.
Is Kimi K3.1 Pro open source?
Not yet in the downloadable sense. Moonshot describes it as an open-weight model and plans to publish its weights by November 4, 2026, but at launch the weights had not been released and the licence had not been published. Until then, you can only use it through Moonshot’s hosted products and API.
How big is Kimi K3.1 Pro?
Moonshot describes it as a 3-trillion-parameter-class model. Artificial Analysis lists 2.8 trillion total parameters, with 104 billion active parameters, in a mixture-of-experts design. It has a 1-million-token context window and accepts text and image input.
What is Kimi K3.1 Pro Highspeed?
It is a faster variant of the same model, released alongside the standard version. Moonshot says it runs at two to three times the speed of Kimi K3.1 Pro with identical model capabilities, aimed at users who need quicker responses.
Why did Moonshot reduce Kimi K3.1 Pro’s hacking ability?
Moonshot said it removed cyber-related data from the training corpus and reinforcement-learning environments so that releasing open weights would not extend the frontier of cyber capability. On its internal benchmark, the filtered model scored 81.0 percent, down from 90.5 percent for an unfiltered version, but still ahead of its predecessor. The results have not been independently verified, and general coding skills were retained.
How much does Kimi K3.1 Pro cost?
Artificial Analysis lists API pricing at $2 per million input tokens and $10 per million output tokens. Hosted use through Kimi.ai depends on Moonshot’s own plans, and self-hosting costs cannot be judged until the weights and licence are published on November 4.