Every major Kimi release in one place, newest first: models, features, Kimi Work and Kimi Code updates, plan changes and API retirements, each with what it means for you.
Kimi K3.1 Pro launched on 23 October 2026 across the Kimi apps, Kimi Work, Kimi Code and the API. On 4 November, Moonshot released its full weights on Hugging Face.
Reads very long documents and big codebases in one go.
Input and output, about a third less than K3.
About 104 billion active for each token.
4th overall at launch, the top open-weight model.
Newest first. Filter by type to see only models, features, app updates or plan and API changes.
Adds global search across projects, plugins, skills and chats, a “Stop and Archive” action, and goal-mode progress shown above the input box.
Kimi Work guide →Moonshot published the full K3.1 Pro weights on Hugging Face, as promised at launch, so developers can run the model themselves.
Kimi K3.1 Pro guide →The CLI can run tools on a remote machine or container. Desktop adds several working folders per project, forking from an earlier turn and one-click diagnostics export.
Kimi Code guide →K2-series API models, including K2.6 and K2.7 Code, are no longer available to new users and retire by the end of December 2026. Existing users should move to K3.1 Pro.
Kimi API guide →Kimi Work now confirms key details with you before it starts a task, so fewer jobs run with the wrong assumptions.
Kimi Work 3.2 →Introduced with K3.1 Pro: create and edit 3D models from text, sketches or images, preview them in the browser and download them for 3D printing.
Kimi 3D Make guide →The new flagship: a further-trained K3 with a 1M-token context, better coding, 3D and finance work, and an API price of $2/$10 per million tokens. Highspeed is 2–3× faster.
Kimi K3.1 Pro guide →Paid plans reopened under new names at $19, $39, $99 and $199 a month. Kimi Code stays included, and its separate weekly cap was dropped.
Kimi pricing →Kimi's coding agent got a desktop app for macOS and Windows, alongside the command-line tool and VS Code extension.
Kimi Code guide →A new coding model in Kimi Code under the existing kimi-for-coding ID, with a 1M-token context and coding close to K3 at lower cost.
Kimi K2.8 guide →The older K2.5 and moonshot-v1 API models were switched off. Requests to them now return errors.
Kimi API guide →Moonshot released the K3 weights and a 47-page technical report covering its architecture and training.
Kimi K3 guide →After heavy demand for K3, Kimi paused new paid subscriptions while it reworked its plans. They reopened in September.
Kimi pricing →A 2.8-trillion-parameter model with 104 billion active, native image and video understanding and a 1M-token context window.
Kimi K3 guide →A coding-focused, thinking-only model for Kimi Code and the API, with open weights.
Kimi K2.7 guide →A desktop app for Windows and Mac that works on your local files, controls your browser through WebBridge and runs up to 300 sub-agents.
Kimi Work guide →A Chrome and Edge extension that lets Kimi's agents click, type and extract data in your real browser, with your logins staying on your machine.
An open-weight model with stronger long coding tasks and Agent Swarm, and a 256K-token context window.
All Kimi models →A cloud version of the OpenClaw agent built into kimi.com, running 24/7 with long-term memory and access to more than 5,000 community skills.
Added native vision, Agent Swarm with up to 100 sub-agents, the Kimi Code command-line agent, and Kimi Docs and Sheets.
All Kimi models →Kimi Slides gained fully editable, designer-style slides.
Create presentations →A reasoning version of K2 that thinks step by step and uses tools many times in a row.
Kimi K2 guide →A small 48B research model testing a faster attention design, later used in K3's architecture.
K2 was updated with a 256K-token context and better coding, and the “OK Computer” agent mode arrived in Kimi.
Kimi K2 guide →A 1-trillion-parameter open-weight model that put Kimi on the map for coding and agent tasks.
Kimi K2 guide →An autonomous research agent in the Kimi app, and a 72B open coding model.
Kimi-Researcher →Open models for speech and audio understanding and for vision.
Kimi-Audio →A reasoning model that brought step-by-step thinking to Kimi.
Kimi began testing a context of about 2 million Chinese characters.
Kimi opened to the public, known for reading very long documents of about 200,000 Chinese characters.
Which models you can use today, and which are on their way out.
| Model | Status | Where you can use it |
|---|---|---|
| Kimi K3.1 Pro | Current flagship | Kimi apps, Kimi Work, Kimi Code, API, open weights |
| Kimi K3.1 Pro Highspeed | Current | Kimi Code (Pro plan and up), API |
| Kimi K3 | Previous flagship | Kimi Code (k3, k3-256k), API |
| Kimi K2.8 Preview | Preview | Kimi Code (kimi-for-coding) |
| Kimi K2.7 Code, K2.6 | Retiring | API: closed to new users 31 Oct 2026, retiring by end of Dec 2026 |
| Kimi K2.5, moonshot-v1 | Retired | Switched off in the API in August 2026 |
Compare them in detail on the Kimi models page.
Long documents have been Kimi's strength since launch. In just over a year, the context window grew from 128K tokens with K2 to 1 million with K3 and K3.1 Pro.
In the Kimi apps, the full 1M context is on Max and Ultra plans. See how to analyze PDFs for tips on long files.
Context window by model, from K2 to K3.1 Pro.
Only changes Moonshot has announced. We don't list rumours.
K2-series API models, including K2.6 and K2.7 Code, are due to retire. Move API projects to K3.1 Pro before then.
K2.8 is still a preview in Kimi Code. Moonshot hasn't given a date for a stable version or API pricing.
Follow @Kimi_Moonshot on X and check the Kimi Code What's New page and the Kimi Work release notes. We update this page with each major release.