OpenModelsHQ

Moonshot AI

Beijing, China · China Desk

Moonshot AI is based in Beijing, China. We track 16 open-weight models from it, and the newest is Kimi-K3.1-Pro, released 2 Nov 2026.

Models tracked
16
LLM 7 · Vision 8 · Speech 1
Newest release
Kimi-K3.1-Pro
2 Nov 2026
Country
China
Releases timeline
16 Moonshot AI releases from 22 Feb 2025 to 2 Nov 2026. The largest is Kimi-K3.1-Pro at 2.78T parameters. Dot area follows total parameters; a ring means the size is not published.
Releases timeline: data
ModelCategoryReleasedParameters
Kimi-K3.1-ProVision2 Nov 20262.78T
Kimi-K3Vision13 Jun 20262.78T
Kimi-K2.7-CodeVision11 Jun 20261.03T
Kimi-K2.6Vision14 Apr 20261.03T
Kimi-K2.5Vision1 Jan 20261.03T
Kimi-K2-ThinkingLLM4 Nov 20251.03T
Kimi-Linear-48B-A3B-InstructLLM30 Oct 202549.1B
Kimi-K2-Instruct-0905LLM3 Sep 20251.03T
Kimi-K2-InstructLLM11 Jul 20251.03T
Kimi-K2-BaseLLM3 Jul 20251.03T
Kimi-VL-A3B-Thinking-2506Vision21 Jun 202516.4B
Kimi-Audio-7B-InstructSpeech25 Apr 20259.8B
Kimi-VL-A3B-ThinkingVision9 Apr 202516.4B
Kimi-VL-A3B-InstructVision9 Apr 202516.4B
Moonlight-16B-A3B-InstructLLM22 Feb 202516B
Moonlight-16B-A3BLLM22 Feb 202516B

Origin

Yang Zhilin, Zhou Xinyu and Wu Yuxin were schoolfriends at Tsinghua University. In March 2023 they founded Moonshot AI in Beijing, according to Wikipedia. Its main products are the Kimi models, which Wikipedia describes as open-weight and available through the web, an API and the Kimi Code command-line agent.

How it got here

The earliest open weights we track from Moonshot are all from 2025. First is Moonlight, a 16B mixture-of-experts model with 3B active parameters. Wikipedia describes a joint Moonshot and UCLA paper that trained a model of that size with the Muon optimiser. The Kimi-VL vision-language models and the Kimi-Audio model came the same year.

Kimi K2 followed in July 2025: a mixture-of-experts model with 1 trillion total parameters, 32 billion active, trained on 15.5 trillion tokens. Wikipedia says it became the most downloaded model on Hugging Face the day after release. An updated K2 version came later.

Kimi Linear, a 48B model with 3B active parameters, came in 2025, and Kimi K2 Thinking, a reasoning update, in November 2025. Kimi K2.5, in January 2026, added native image and video input through a vision encoder called MoonViT. Further K2 releases continued through 2026, and then Kimi K3, which Wikipedia puts at 2.8 trillion parameters.

Licensing pattern

Moonshot started on plain MIT and has moved to terms with conditions.

Moonlight and Kimi Linear use the standard MIT licence, according to their model cards.

Kimi K2 uses a modified MIT licence: MIT plus one condition. A commercial product with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, must display “Kimi K2” on its user interface.

Kimi K3 has a custom Kimi K3 License that keeps MIT-style permissions and adds two conditions. A business offering the model as a service to third parties must make a separate agreement with Moonshot if its revenue, with its affiliates, exceeds 20 million US dollars over 12 months. Wikipedia reports that this involves revenue sharing of up to 30%. And products above 100 million monthly active users or 20 million US dollars in monthly revenue must display “Kimi K3”. Internal use, and use through Moonshot’s official products or certified inference partners, is excluded from both conditions.

A company that hosts K3 as a paid inference service should read the licence before it starts.

Release lineage

Newest first
  1. Kimi-K3.1-Pro2.78T
  2. Kimi-K32.78T
  3. Kimi-K2.7-Code1.03T
  4. Kimi-K2.61.03T
  5. Kimi-K2.51.03T
  6. Kimi-K2-Thinking1.03T
  7. Kimi-Linear-48B-A3B-Instruct49.1B
  8. Kimi-K2-Instruct-09051.03T
  9. Kimi-K2-Instruct1.03T
  10. Kimi-K2-Base1.03T
  11. Kimi-VL-A3B-Thinking-250616.4B
  12. Kimi-Audio-7B-Instruct9.8B

Largest first. The last line on each tile is the memory of the smallest device that fits the model at 4-bit.