Kimi

Kimi is Moonshot AI's chat assistant and family of open-weight models. The newest, Kimi K3.1 Pro (Nov 2, 2026), is a 2.8-trillion-parameter agentic model with a 1-million-token context, free to download under a license that adds terms only for large model-as-a-service companies.

Kimi is the AI assistant at kimi.com made by Moonshot AI of Beijing, and the name of its open-weight models. The newest, Kimi K3.1 Pro, released on Nov 2, 2026, is a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active, a 1-million-token context and native image understanding, built for long coding and agent tasks. Its weights are free to download, and the Kimi API is pay as you go (as of Nov 6, 2026).

Kimi's strength is agentic work: long coding sessions, research and building documents and slides with tools, at open-weight terms. Pick it to run a frontier-class open model yourself or as a cheaper engine for coding agents such as Claude Code and Codex, which Moonshot documents. For an everyday assistant with the widest features, ChatGPT, Claude or Gemini fit better. See Best AI Chatbot and Best AI Coding Agent.

Quick facts

Maker	Moonshot AI, Beijing
Get it	kimi.com, mobile apps, Kimi API, Kimi Code
Newest model	Kimi K3.1 Pro (Nov 2, 2026)
Size	2.8 trillion parameters, 104 billion active
Context	1,048,576 tokens for K3 models (Kimi API pricing)
Open weights	Yes, under the Kimi K3.1 Pro License, a permissive license with one revenue condition
API price	Pay as you go, no subscription; batch jobs at 60% of the standard price
Best for	Long coding and agent tasks, open weights at frontier scale

Models

Model	Released	Notes
Kimi K3.1 Pro	Nov 2, 2026	Post-trained from K3; Moonshot's most capable model
Kimi K3	Jun 2026	2.8 trillion parameters; Moonshot calls it the first open 3T-class model
Kimi K2.7 Code	Jun 2026	Coding model
Kimi K2.6	Apr 2026
Kimi K2 Thinking	Nov 2025	Reasoning version of K2
Kimi K2	Jul 2025	1 trillion parameters, 32 billion active, modified MIT license

Dates are when Moonshot posted the weights on Hugging Face. Moonshot says the K2 series closed to new API users on Oct 31, 2026 and will be retired by the end of Dec 2026.

Features

Agentic coding. K3.1 Pro is tuned for long engineering sessions with little supervision. Moonshot tests it inside Claude Code and its own Kimi Code, and the Kimi API docs explain how to plug it into Codex, OpenCode and Hermes Agent through OpenAI- and Anthropic-compatible endpoints.

Reasoning effort. Developers choose low, high or max thinking; max is the default, per the model card.

Multimodal. K3 and K3.1 Pro read images and video natively, and Moonshot reports results on 3D modeling, slide decks and design tasks.

Knowledge work. The Kimi app highlights complex coding projects, 3D models and in-depth financial research for K3.1 Pro, as its home page says.

0:00 Agent Swarm. With Kimi K2.6 in Apr 2026, Moonshot introduced Agent Swarm, which splits one request among many subagents working in parallel; its launch clip shows one prompt turned into reports, slide decks and spreadsheets.

1:45 In an Apr 22, 2026 video, Caleb Writes Code used Agent Swarm with K2.6 to gather about 1,500 US data centers into one Excel file: after a few minutes spent writing a detailed brief, Kimi launched waves of subagents, each researching its own share and reporting back to a main agent.

License

The Kimi K3.1 Pro License lets anyone use, modify, fine-tune and sell the model. A company that runs a model-as-a-service business, letting others call the model through an API, must sign a separate agreement with Moonshot once its group's revenue passes $20 million over any 12 months. End-user products that only build the model into their own features are not counted as model-as-a-service. The earlier Kimi K2 used a modified MIT license.

API price

The Kimi API bills input and output tokens with automatic context caching and no subscription, and batch jobs cost 60% of the standard price. The per-token prices for K3.1 Pro did not load in a form that could be read on Nov 6, 2026, so check Moonshot's pricing page for the current figures. Kimi membership and Kimi Code are separate subscriptions from the API.

Kimi vs the alternatives

DeepSeek has the lowest published API prices among the Chinese labs and an MIT license with no revenue condition. GLM sells a dedicated coding subscription from $18 a month; Zhipu's own table puts GLM-5.3 (88.2) and Kimi K3 (88.3) within a point of each other on Terminal Bench 2.1. Qwen has the widest range of sizes, including small Apache 2.0 models. Claude remains the reference for coding agents, and Moonshot compares K3.1 Pro against Claude Opus 5.5 in its own benchmarks. Pick Kimi for the largest open model aimed at agents.

Frequently asked questions

Is Kimi free? Yes. The Kimi assistant at kimi.com has free use, and Kimi's model weights are free to download; the API and the Kimi membership and Kimi Code plans are paid.

What is Kimi K3.1 Pro? Kimi K3.1 Pro is Moonshot AI's newest open-weight model, released Nov 2, 2026: 2.8 trillion parameters with 104 billion active, a 1-million-token context and native vision, built for long coding and agent work.

Is Kimi K2 still available? Kimi K2 models remain open weight on Hugging Face, but Moonshot's API stopped offering them to new users on Oct 31, 2026 and plans to retire them by the end of Dec 2026.

Who makes Kimi? Kimi is made by Moonshot AI, an AI company based in Beijing.

Can I use Kimi in Claude Code? Yes. Moonshot's API offers Anthropic- and OpenAI-compatible endpoints, and its documentation covers Claude Code, Codex, OpenCode and its own Kimi Code CLI.