Kimi K3.1 Pro: Moonshot AI’s 3-Trillion-Parameter Open-Weight Model

Key Takeaways

• Moonshot AI has unveiled Kimi K3.1 Pro, a 3-trillion-parameter open-weight model with a 1-million-token context window, alongside a faster variant called Kimi K3.1 Pro Highspeed.
• Both models are available now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API, according to a company statement reported by ANI.
• Moonshot AI plans to release the full model weights by November 4, 2026, which is the step that would turn an API-accessible model into one developers can download and run themselves.
• On Terminal-Bench 4.0, the company reports a score of 65.7 percent, ahead of GPT-6 Astra at 58.2 percent and just behind Claude Opus 5.5 at 66.4 percent, and these are company-reported figures, not independently verified results.
• The company positions the model for coding across full projects, 3D creation, financial research and other agentic work, with improvements to reinforcement learning, long-context handling and training efficiency.
• For most readers, the sensible move is to try the hosted version on a real task now and wait for the weights release, with its licence terms, before planning any self-hosted deployment.

What Moonshot AI announced

Chinese AI company Moonshot AI has unveiled Kimi K3.1 Pro, the first model in its new K3.1 series. According to a company statement carried by ANI, it is the company’s most capable 3T-class open-weight model to date, with a 1-million-token context window. The announcement also introduced Kimi K3.1 Pro Highspeed, a faster variant of the same model, and both versions are available now across Kimi.ai, Kimi Work, Kimi Code and the Kimi API.

Moonshot AI says the model delivers substantial improvements across coding, 3D creation, financial research and other agentic capabilities. Agentic work here means tasks where a model does more than answer a question, such as carrying out multi-step work in a codebase, gathering and checking information, or completing a workflow from start to finish. The company also says it optimised its reinforcement learning framework, long-context processing, training efficiency and infrastructure, with the aim of improving performance across hundreds of application areas, including coding, autonomous AI tasks and knowledge-based work.

What 3 trillion parameters and 1 million tokens actually mean

Two numbers dominate the announcement, and both deserve a plain explanation, because neither one on its own tells a reader whether a model is good.

The parameter count

A parameter is, in simple terms, one of the adjustable values a model learns during training. Three trillion of them puts Kimi K3.1 Pro in the largest class of models currently discussed in public, which is why Moonshot AI describes it as a 3T-class model. A larger parameter count can support more knowledge and more capable reasoning, but it is not a score. How the model was trained, what data it learned from, and how it was tuned matter at least as much, and two models with similar sizes can behave very differently on the same task.

The context window

The 1-million-token context window is the amount of text and other input the model can consider in a single session. In practical terms, that is enough to hold a large codebase, a long set of documents, or a lengthy working history at once, rather than forcing a user to split work into small pieces. For coding, that means the model can in principle see related files, tests and documentation together. For research, it means long reports and source material can sit in the same conversation as the questions about them.

The benchmark numbers, with a caution

The most quoted figure from the announcement is the Terminal-Bench 4.0 result. Terminal-Bench is a benchmark that evaluates coding agents on real-world tasks carried out in a terminal, which makes it a reasonable proxy for practical coding ability, though still only one test among many. Moonshot AI reports that Kimi K3.1 Pro scored 65.7 percent, surpassing GPT-6 Astra at 58.2 percent and coming close to Claude Opus 5.5 at 66.4 percent.

Those figures should be read with care. They are company-reported results, published as part of a product announcement, and they have not been independently verified in the reporting reviewed for this article. Benchmarks also depend on the exact test setup, prompting and tooling used, so small gaps can shift between evaluations. The honest reading is narrower than the headline suggests. Kimi K3.1 Pro appears competitive with leading proprietary models on this particular coding benchmark, according to its maker, and that is a meaningful claim for an open-weight model. It is not proof that it beats those models in general, and readers should not treat a single benchmark as a verdict.

For a broader look at how another frontier model family is positioned, see our guide to Google Gemini 3.

Kimi K3.1 Pro or Highspeed: which one to try

Moonshot AI has shipped two versions, and the choice between them is a practical one. Kimi K3.1 Pro is the standard model, intended as the full-capability option. Kimi K3.1 Pro Highspeed is the faster variant, aimed at work where response speed matters more than squeezing out the last bit of quality on a hard task.

Most readers should start with the standard model for anything where accuracy matters, such as reviewing code, working through a financial document, or building a project end to end. The Highspeed variant makes more sense for interactive work, such as quick fixes, drafting, or iterating rapidly, where waiting longer for each response would break the flow. If a task is important, it is worth running it on the standard model at least once and comparing the result, rather than assuming the faster answer is equally good.

Availability is straightforward. Both versions are accessible through Kimi.ai for chat-style use, Kimi Work for document and research-style work, Kimi Code for coding, and the Kimi API for developers building their own tools. That range means a reader can try the model without writing code, while a developer can test it programmatically in the same week.

What the November 4 weights release would change

The most consequential date in the announcement is November 4, 2026, when Moonshot AI plans to release the full model weights. Weights are the learned values that make up the trained model, and releasing them is what allows developers and researchers to download, inspect, adapt and run a model on their own infrastructure, subject to the licence that accompanies the release.

Until that release happens, Kimi K3.1 Pro is best described as API-accessible rather than fully open in the hands-on sense. Readers can use it through Moonshot AI’s products, but they cannot yet host it themselves or fine-tune it on their own data in the way a weights release would allow. That distinction matters for teams with strict data rules, for researchers who want to study the model directly, and for developers in regions or organisations where sending code or documents to an external API is not acceptable.

A practical way to evaluate it

For developers, the strongest test is a task taken from real work. Pick a small but genuine job, such as fixing a bug in an existing repository, adding a feature with tests, or refactoring a messy module, and run it through Kimi Code. Keep the original code, the prompt and the output, then check the result by running the tests and reviewing the changes line by line. A model that produces plausible-looking code that fails tests or introduces subtle errors is less useful than a slower model that gets the details right. For the production habits that matter once a new tool enters an existing workflow, see our guide to video editing tools and a scalable content workflow.

In all three cases, the same rule applies. Judge the model on your own task, with your own checks, and keep the benchmark numbers in their proper place, as a reason to try it rather than a conclusion about it.

What is still unconfirmed

Several important details were not included in the company statement as reported, and readers should not fill those gaps with assumptions. Pricing for the Kimi API was not part of the reported announcement summary, so anyone planning a budget should check the current Kimi API documentation rather than relying on second-hand figures. The licence for the forthcoming weights release is likewise unconfirmed until Moonshot AI publishes it. Independent benchmark results, beyond the company’s own Terminal-Bench figures, have not yet appeared in the reporting reviewed here, and the model’s behaviour on languages, domains and tasks outside the highlighted areas remains to be tested by users.

That is normal for a launch announcement. The confirmed facts are the model itself, its size class and context window, its availability across Moonshot AI’s products, the Highspeed variant, the planned November 4 weights release, and the company-reported benchmark scores. Everything else is either a company claim to be tested or a detail to await.

Bottom line

Kimi K3.1 Pro is a significant release because it brings a very large, long-context model toward the open-weight world, with coding and agentic work as its clearest strengths. The reported Terminal-Bench result suggests it can stand near leading proprietary models on practical coding tasks, but that suggestion comes from the company itself and needs independent testing. The hosted versions are available now for anyone who wants to try them, and the real test of Moonshot AI’s openness will come on November 4, when the full weights are due. Until then, try it on real work, verify the outputs, and judge it by what it does for you.

Sources: the company statement as reported by ANI in The Economic Times and The Tribune.