No longer served by Moonshot's API, so this is now only for developers willing to host a trillion-parameter model themselves.
Kenning's take
What it does
Kimi K2 is a mixture-of-experts language model from Moonshot AI with 1 trillion total parameters, 32 billion of them active per token, and a 128K context window. It came as a Base model for fine-tuning and an Instruct model for chat and agent tasks. Moonshot's API stopped serving the K2 series on 25 May 2026, so it is now used through the open weights on Hugging Face.
- 1 trillion total parameters, 32 billion active
- 128K token context
- Base and Instruct versions
- Trained on 15.5 trillion tokens with the MuonClip optimiser
- Modified MIT licence for code and weights
Who it suits. Developers and researchers who want to run or fine-tune a large open model for coding and agent work.
Key facts
- Category
- AI models
- Part of
- Kimi
- Made by
- Moonshot AI
- Based in
- China
- Released
- 11 July 2025
- How to get it
- open weights on Hugging Face under a Modified MIT licence; Moonshot's own API discontinued the K2 series on 25 May 2026
- Works on
- self-hosted, Hugging Face
- Free plan
- Weights free to download under the Modified MIT licence
FAQ
When was Kimi K2 released?
Its Hugging Face repository was published on 11 July 2025.
Is Kimi K2 still available?
The weights remain on Hugging Face under a Modified MIT licence. Moonshot's API stopped supporting the K2 series on 25 May 2026 and points users to Kimi K3.
How big is Kimi K2?
1 trillion total parameters with 32 billion active per token, and a 128K context window.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools

Kimi K2 Thinking
Reasoning version of Kimi K2 that thinks step by step between hundreds of tool calls

Kimi K2.5
Moonshot AI's open multimodal model that added image and video input and agent swarms to Kimi

Kimi K2.6
Open Moonshot AI model for long coding sessions and swarms of up to 300 sub-agents

Kimi K2.7 Code
Moonshot AI's open coding model, built on K2.6 and using about 30% fewer thinking tokens