Kimi
A Beijing lab's open-weight model family that made its name on long context — and now coordinates hundreds of AI agents working in parallel.
What is Kimi?
Kimi is built by Moonshot AI, a Beijing startup founded in 2023 by former ByteDance employees and backed by Alibaba and HongShan (formerly Sequoia China). Its founder, NLP researcher Zhilin Yang, bet early on two things that took a while for the rest of the industry to catch up on: very long context windows, and models that could act rather than just answer. The first version of Kimi made headlines for a record-setting context window at launch; two years later, that same bet has turned into a full model family and a genuinely competitive position against much better-funded Western labs.
The current flagship, Kimi K3, released July 16, 2026, is a 2.8-trillion-parameter model with native vision and a 1-million-token context window that stays in "thinking" mode by default. Independent testing has placed it second among dozens of tracked models — behind only OpenAI's top-tier GPT-5.6 Sol, which happens to be restricted in some markets. That makes Kimi K3 arguably the strongest model generally accessible without a VPN in regions where the leading Western models aren't available, which is a real and unusual position for an AI product to be in.
Benchmark ranking comes from independent third-party testing reported in mid-2026 and reflects a snapshot in a fast-moving field. Model lineup and figures are current as of July 2026 and change as Moonshot ships new versions.
Key features
Very long context
Handle entire books, codebases, or document sets in a single conversation without losing earlier detail.
Agent Swarm
Coordinates up to 300 parallel sub-agents to decompose and complete a large task autonomously.
Kimi Work
A desktop agent that carries out tasks directly on your computer, not just inside a chat window.
Kimi Code
A coding-focused product built on the K2.7 Code model, for end-to-end programming tasks.
Multimodal input
Understands text, images and video, not just written prompts.
Open-weight releases
Several Kimi models can be downloaded and self-hosted under a modified MIT license.
Pricing
Free chat app
- Access to Kimi's consumer chat app
- Long-context conversation included
- No monthly plan to compare or upgrade
API — K2.6
- General-purpose flagship for most tasks
- 256K-token context window
- Open-weight, self-hostable option available
API — K3
- Current flagship, top-tier benchmark results
- 1M-token context window
- Always-on reasoning by default
There is no paid consumer subscription — the chat app is free, and Moonshot monetizes through per-token API usage. Cache-hit input pricing is substantially cheaper than shown. Prices reflect Moonshot AI's official API pricing as of July 2026 and may change.
Available models
Integrations & platforms
Pros, cons & best for
Pros
- Genuinely class-leading context window (up to 1M tokens)
- Competitive with top closed models at a fraction of the API cost
- Several models are open-weight and self-hostable
Cons
- Rapid model turnover makes it hard to know which version you're using
- Some benchmark claims come from Moonshot's own test suites, not independent ones
- Chinese data hosting is a real consideration for sensitive use cases
Best for
- Very long documents, codebases, or research materials
- Developers wanting a cheap, capable, self-hostable model
- Multi-step agentic workflows via Agent Swarm
Take a look inside
Alternatives
For another cheap, open-weight option from a different region:
Our verdict
Kimi's bet on long context and agentic capability has paid off — K3 genuinely competes with the best closed models available anywhere, and the open-weight releases give developers a real, self-hostable alternative at a fraction of the API cost. The catch is pace and provenance: Moonshot ships new versions fast enough that keeping track of which model you're actually talking to takes effort, and some of the more impressive benchmark claims haven't yet been independently verified. For anyone working with very long documents or codebases, or wanting a capable model they can run themselves, Kimi is one of the strongest options on this page.
FAQ
Is Kimi free to use?
Yes, the consumer chat app is free with no subscription. Moonshot AI charges separately for API usage, billed per token.
Who makes Kimi?
Moonshot AI, a Beijing-based startup founded in 2023 by former ByteDance employees, backed by Alibaba and HongShan (formerly Sequoia China).
What is Agent Swarm?
A mode that coordinates up to 300 parallel sub-agents to break a large task into pieces and complete it autonomously, rather than handling everything in a single linear conversation.
Can I self-host a Kimi model?
Yes, several versions, including K2.6, are released under a modified MIT license and can be downloaded and run independently of Moonshot's own servers.
How large a document can Kimi handle?
The current flagship, Kimi K3, supports a 1-million-token context window — enough for very long documents, books, or codebases in a single conversation.
Is Kimi's data hosted in China?
Yes, as a Chinese company, Moonshot AI is subject to Chinese data regulations — worth factoring in if you're handling sensitive information.