Chinese artificial intelligence company Moonshot AI has launched Kimi K3.1 Pro, an upgraded large language model designed for advanced coding, financial research, 3D creation and autonomous AI-agent workflows. Announced on October 24, 2026, the model features a 1-million-token context window and is available through Kimi’s website, mobile apps, Kimi Work, Kimi Code and the Kimi API. The company has also introduced Kimi K3.1 Pro Highspeed, a variant it says runs two to three times faster than the standard model.
The launch strengthens Moonshot AI’s position in the competition to develop powerful AI models that can eventually be distributed with open weights. The company says Kimi K3.1 Pro delivers competitive results on several coding and knowledge-work benchmarks, while targeted training changes reduce its performance on certain cybersecurity tests. Full model weights are scheduled for release by November 4, 2026, although they had not yet been published at launch. This distinction matters because access to a hosted model is not the same as downloading its weights to run or modify it independently.
Key Takeaways
- New flagship model: Kimi K3.1 Pro is an upgraded version of Moonshot AI’s Kimi K3.
- 3-trillion-class architecture: The model belongs to the 3T-class family; reporting on the architecture puts it at approximately 2.8 trillion total parameters, with around 104 billion active at a time.
- Long context: It supports a 1-million-token context window for processing large amounts of information.
- Faster variant: Kimi K3.1 Pro Highspeed is claimed to deliver two to three times the speed of the standard model.
- Coding and 3D design: The company reports strong results in coding-agent and 3D-modeling evaluations.
- Cybersecurity safeguards: Moonshot says it removed cyber-related training data and filtered reinforcement-learning environments to reduce offensive cyber capabilities.
- Open weights planned: The company has set November 4, 2026, as the target for releasing the full model weights.
What Is Kimi K3.1 Pro?
Kimi K3.1 Pro is a post-trained upgrade of Moonshot AI’s Kimi K3 model. Rather than introducing an entirely separate architecture, the company says it invested substantially more computing resources in post-training and improved the systems used for reinforcement learning, long-context inference and evaluation.
Post-training helps refine a model’s behavior after its initial training. It can improve how a model follows instructions, uses tools, solves multi-step problems and performs on specialized tasks.
Moonshot says the new version improves across coding, autonomous-agent workflows, financial analysis and 3D creation. The company also introduced Kimi K3.1 Pro Highspeed for applications where faster responses are a priority.
The model is available through the company’s hosted services, but the planned open-weight release is a separate milestone. Open weights allow developers to download a model’s learned parameters and potentially run it on their own infrastructure, subject to the eventual license and technical requirements.
Understanding the 3-Trillion-Parameter Architecture
The headline’s 3-trillion figure describes the model’s approximate parameter class rather than a claim that every parameter is used for every request.
Kimi’s K3 family uses a mixture-of-experts architecture. Instead of activating the entire network for every input, the system routes each token through selected expert components. This allows a model to have a very large total parameter count while using a smaller active subset during inference.
| Model characteristic | Reported specification |
|---|---|
| Model family | Kimi K3.1 |
| Total parameters | Approximately 2.8 trillion, or 3T-class |
| Active parameters | Approximately 104 billion |
| Context window | 1 million tokens |
| Architecture | Mixture of experts |
| Faster version | Kimi K3.1 Pro Highspeed |
| Planned weights release | November 4, 2026 |
The distinction between total and active parameters is important. A larger total parameter count does not automatically mean that a model will outperform a smaller model on every task. Performance also depends on training quality, architecture, inference settings, tool use and the evaluation being measured.
The 1-million-token context window is another notable feature. It can allow the model to process large collections of documents, long codebases or extensive research material within a single context, although practical performance still depends on how effectively the model retrieves and reasons over information.
Coding Performance and AI-Agent Capabilities
Coding is one of Moonshot’s main selling points for Kimi K3.1 Pro. The company says the model can handle tasks ranging from small fixes in existing codebases to full-project development.
On Terminal-Bench 4.0, a benchmark that tests coding agents on realistic terminal tasks, Moonshot reported a score of 65.7%. Its published comparison listed GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4%.
Kimi also reported a score of 61.2% on its internal Kimi Code Bench 3.0 Long-horizon evaluation, compared with 55.95% for the previous K3 model.
These results suggest improvements in multi-step coding and tool use, but benchmark comparisons require context. Some results come from company-run tests, and evaluation harnesses or model settings can differ. The figures should not be interpreted as proof that Kimi K3.1 Pro is universally better than competing models.
The broader goal is to make the model useful for agentic coding: systems that can inspect a repository, edit files, run tests, diagnose failures and continue iterating until a task is complete. Such workflows can reduce manual work, but they still require code review, testing and safeguards before deployment.
Stronger 3D Creation and Financial Research
Moonshot is also positioning Kimi K3.1 Pro as a model for work beyond conventional chat and coding.
In 3D creation, the company reported a score of 92.7% on BenchCAD with tools. This evaluation measures the ability to reconstruct 3D objects from multiple views by generating CAD code. Moonshot’s reported comparison placed the model near more established frontier systems.
The capability could help designers and developers generate parametric models, build interactive 3D scenes and create assets for games or product visualization. The output still needs technical validation, especially when designs must meet manufacturing tolerances or other precise requirements.
For financial research, Moonshot reported a score of 64.4 on OfficeQA Pro v2, a benchmark involving complex financial-document analysis. It also reported improvements over K3 on internal financial modeling and research evaluations.
These capabilities could support document analysis, spreadsheet workflows and investment research. However, model-generated financial analysis should not be treated as verified advice: source checking, numerical validation and human judgment remain essential.
How Moonshot Is Addressing Cybersecurity Risks
One of the most consequential parts of the announcement concerns the model’s ability to discover and exploit software vulnerabilities.
Moonshot says it identified cyber-related material in its training corpus and reinforcement-learning environments, then removed or filtered that material for the Kimi K3.1 Pro open-weight release. The company says this reduced offensive cyber capability while preserving general-purpose coding skills.
On its internal xdev-bench-hard evaluation, Moonshot reported that a checkpoint without these risk-reduction measures solved 90.5% of problems, compared with 81.0% for the final Kimi K3.1 Pro model.
These figures are company-reported and have not been independently verified in the sources reviewed. They also show a reduction in capability, not its elimination. Strong general coding and debugging skills can still transfer to some cybersecurity tasks.
The issue is especially relevant for open-weight models. Once weights are widely distributed, the developer has less control over how the model is deployed than it would with a hosted API. Moonshot argues that open models can help defenders audit software and find weaknesses, but the same capabilities can have dual-use implications.
Availability, Access and the Planned Open-Weight Release
Kimi K3.1 Pro and its Highspeed variant are available through Kimi’s hosted products, including Kimi.com, Kimi Work, Kimi Code and the Kimi API.
The full model weights are scheduled for release by November 4, 2026. Until that release occurs, developers should not assume that downloadable weights or final license terms are already available.
For organizations, the choice between hosted access and self-hosting will depend on latency, privacy, infrastructure costs, operational expertise and licensing. Hosted services can simplify deployment, while downloadable weights may offer greater control over infrastructure and customization—but typically require substantial compute resources and technical maintenance.
The Bigger Picture
Kimi K3.1 Pro arrives as Chinese AI companies continue competing with major U.S. developers on coding, reasoning and autonomous-agent tasks. Its long context window, mixture-of-experts architecture and reported benchmark results illustrate how model competition is increasingly about practical work: building software, handling complex documents and operating tools across many steps.
The planned open-weight release could make the model more accessible to developers and businesses that want to run AI outside a proprietary hosted service. At the same time, Moonshot’s decision to reduce certain cybersecurity capabilities highlights the challenge of balancing openness with misuse prevention. The ultimate impact will depend on independently verified performance, the final license and how the released model behaves in real-world deployments.
FAQs
1. What is Kimi K3.1 Pro?
Kimi K3.1 Pro is Moonshot AI’s upgraded large language model for coding, financial research, 3D creation and agentic workflows. It features a 1-million-token context window.
2. Does Kimi K3.1 Pro have exactly 3 trillion parameters?
The model is described as 3T-class. Reporting on the K3 architecture gives an approximate total of 2.8 trillion parameters and around 104 billion active parameters. The total count and active count are different measures.
3. How much faster is Kimi K3.1 Pro Highspeed?
Moonshot says the Highspeed variant delivers two to three times the speed of the standard model. Actual latency can vary by workload, request size and service conditions.
4. Is Kimi K3.1 Pro open source?
The model is available through hosted products and an API, while the company has promised a full open-weight release by November 4, 2026. Open weights are not automatically the same as open-source software; the eventual license will determine how the model can be used, modified and redistributed.
5. Is Kimi K3.1 Pro safe for cybersecurity work?
Moonshot says it reduced the model’s offensive cyber capabilities through training-data removal and reinforcement-learning filters. This does not eliminate risk. Organizations should use access controls and authorized testing environments when applying AI to vulnerability research.
Looking Ahead
The next major milestone is the planned November 4 release of the full model weights. Developers will be able to assess the practical requirements for running the model, examine the license and test whether the reported coding and reasoning capabilities translate to their own workloads.
Independent evaluations will also be important, particularly for cybersecurity safeguards and long-context performance. If Kimi K3.1 Pro combines competitive task performance with a usable license and manageable deployment costs, it could become a significant option for teams seeking alternatives to closed, hosted AI models.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
