Abstract AI development workspace with code, documents, and a 3D model

Kimi K3.1 Pro is aimed at coding, agent workflows, and long-context tasks.

Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, making the new AI model available through Kimi apps, Kimi Code, Kimi Work, and the Kimi API. The Chinese AI company says the model improves on Kimi K3 for coding, AI-agent tasks, 3D creation, and financial research; it also plans to release the full model weights by November 4, 2026.

Key takeaways

  • Kimi K3.1 Pro is available now as the API model ID kimi-k3.1-pro, with a faster kimi-k3.1-pro-highspeed option.
  • Moonshot describes K3.1 Pro as a 3T-class model with a 1 million-token context window and native image and video input support.
  • The Kimi API supports OpenAI-compatible Chat Completions and Responses APIs, plus an Anthropic-compatible Messages API.
  • Moonshot says full K3.1 Pro weights are scheduled for release by November 4, 2026; the license terms were not specified in the announcement.
  • Moonshot says it removed cyber-related training data to reduce misuse risk before releasing the model weights.

What did Moonshot AI announce?

Moonshot AI introduced Kimi K3.1 Pro as the first model in its K3.1 series. In its official technical announcement, the company called it its most capable 3T-class open-weight model to date and said it has a 1 million-token context window.

For beginners, a context window is the amount of text and other input an AI model can consider in one request or conversation. A 1 million-token window is intended for work such as analyzing large codebases, lengthy document collections, or extended agent workflows. A token is a small unit of text; Kimi’s documentation estimates that one token is roughly three to four English characters.

The release is an update to Kimi K3, which Moonshot released in July 2026. Reporting from LinkedIn and The Information described K3.1 Pro as an upgrade focused on coding and agentic capabilities, with benchmark claims also covering finance and 3D design.

How does Kimi K3.1 Pro work for developers?

Diagram of an AI agent using documents and tools to produce output

AI agents can combine model reasoning with multi-step tool use.

Kimi K3.1 Pro is a post-trained version of Kimi K3. Moonshot says it used substantially more post-training compute and reinforcement learning, or RL, to improve the model across coding, tool use, long-context work, and other downstream tasks.

An AI agent is software that can take a multi-step task, use tools, and act on intermediate results instead of only returning a single chat response. Kimi’s API documentation lists support for tool calling, streaming output, JSON mode for structured responses, multi-turn chat, and configurable reasoning effort.

The API offers three integration formats. Developers can use OpenAI-compatible Chat Completions and Responses APIs, or an Anthropic-compatible Messages API. That compatibility can reduce migration work for projects already using those SDK formats, although teams should still test response formats, tool behavior, limits, and safety handling before production use.

OptionModel IDWhat Moonshot says
Kimi K3.1 Prokimi-k3.1-proFlagship K3.1 model with a 1 million-token context window.
Kimi K3.1 Pro Highspeedkimi-k3.1-pro-highspeedSame capabilities at about two to three times the speed, according to Moonshot.

What coding and benchmark results did Moonshot report?

Moonshot reported a 65.7% score for K3.1 Pro on Terminal-Bench 4.0, a benchmark for coding agents performing resource-constrained terminal tasks. The company compared that result with 58.2% for OpenAI’s GPT-6 Astra and 66.4% for Anthropic’s Claude Opus 5.5.

Those figures should be read as vendor-reported benchmark results, not as a guarantee that one model will be better for every codebase or task. The Next Web noted that the figures had not been independently checked. Developers evaluating K3.1 Pro should use representative repositories, tests, security controls, and cost limits rather than relying only on leaderboard results.

Moonshot also reported a 61.2% result on its internal Kimi Code Bench 3.0 Long-horizon benchmark, compared with 55.95% for K3. Because that is an internal benchmark, it is particularly important to treat the comparison as a company claim rather than an independent measurement.

How did Moonshot address cybersecurity risks?

Abstract security filter protecting a software environment from AI misuse

Moonshot says it removed cyber-related training data before the planned weight release.

Moonshot said it identified and removed cyber-related data from both mid-training and reinforcement-learning stages for the K3.1 Pro release. The company said the measure was designed to reduce the risk that an openly released model could be misused for offensive cybersecurity work while retaining general coding usefulness.

In Moonshot’s internal xdev-bench-hard evaluation, the released K3.1 Pro checkpoint achieved an 81.0% pass rate, while a version without the cited cyber risk-reduction measures reached 90.5%. The company also said the released model improved over K3, which scored 33.3% in that evaluation.

The approach does not mean developers should rely on the model as a security control. AI-generated code and vulnerability analysis require human review, controlled test environments, and standard secure-development practices.

When is Kimi K3.1 Pro available?

Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are available now through Kimi.ai, Kimi mobile apps, Kimi Work, Kimi Code, and the Kimi API. Kimi Work requires version 3.1.0 or later on Windows or Apple silicon Macs, according to Moonshot.

For API users, K3.1 Pro requires a successful account top-up of at least $1 to unlock access. Moonshot describes API billing as pay-as-you-go and says usage is billed for input, output, and context-cache activity. The company’s public documentation accessed for this report did not provide a static per-token rate table.

Moonshot says the full model weights will be released by November 4, 2026. “Open-weight” means the parameter files used by a model are made available for others to download and run, but it does not automatically mean the model is open source. Moonshot had not specified the K3.1 Pro weight license in the materials reviewed for this report.

What should beginners know before trying the API?

Kimi’s quickstart documentation recommends creating an API key, storing it in an environment variable, choosing a model, and then calling the service through HTTP or a compatible SDK. The documentation lists Python 3.9 or later, Node.js 18 or later, and OpenAI SDK 1.0.0 or later for its starter examples.

K3.1 Pro has reasoning enabled by default. Developers can set reasoning_effort to low, high, or max, with max as the documented default. Moonshot also documents fixed values for several sampling settings, so applications should follow the model-specific parameter reference rather than assuming all OpenAI-compatible parameters can be changed.

For early experiments, a small prompt, a test API key with spending controls, and an evaluation set with known answers are safer starting points than connecting an agent directly to production systems. Teams should also avoid embedding API keys in source code or client-side applications.

Frequently asked questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI’s flagship model in the K3.1 series. It is a post-trained update to Kimi K3 that Moonshot says improves coding, agentic work, 3D creation, and financial research.

Is Kimi K3.1 Pro available now?

Yes. As of October 24, 2026, it is available through Kimi products and the Kimi API. Moonshot says the full model weights are planned for release by November 4, 2026.

Which API model IDs should developers use?

The standard model ID is kimi-k3.1-pro. The faster option is kimi-k3.1-pro-highspeed.

Is the Kimi API compatible with OpenAI tools?

Kimi documents OpenAI-compatible Chat Completions and Responses APIs, and says developers can use the OpenAI SDK. It also provides an Anthropic-compatible Messages API.

Are Kimi K3.1 Pro benchmark results independently verified?

No. Moonshot published the cited benchmark results, and The Next Web reported that the figures had not been independently verified. Developers should run their own evaluations before making production decisions.

Sources