Moonshot AI, a Chinese AI Lab backed by Ali Baba and Tencent, has just released a breakthrough open-source model: Kimi K2 Thinking, a trillion-parameter Mixture-of-Experts model built for deep reasoning.
It is the first time ever an open-source model matches the State Of The Art in reasoning tasks, which probably makes Open AI and Anthropic shiver. The model scores an impressive 44.9 with tools at Humanity Last Exam , above GPT 5 and Claude Sonnet 4.5 Thinking models, and matches the current SOTA models on a number of tasks.
The breakthrough is of the same scale as the release of DeepSeek. Once again a Chinese lab has pushed further the limits of efficiency, using INT4 quantization to reach impressively low inference latency and GPU usage while maintaining accuracy.
The model has been optimized for chain-of-thought reasoning and tool calling, autonomously using calls for hundreds of steps, making it ideal to build efficient agentic frameworks.
The model is widely available, already integrated in numerous platforms, including on HuggingFace.
You can try it now:
If you want to learn more about the base model Kimi K2, on which the reasoning model relies, check out Moonshot AI’s Open Agentic Intelligence paper on arxiv.
