Moonshot AI Releases Kimi-K2 Reasoning Model
The new Mixture-of-Experts model is designed for complex tasks but arrives in a custom compressed format with a restrictive license.
Chinese AI startup Moonshot AI has released Kimi-K2-Thinking, a new open-weight model aimed at tackling complex reasoning problems. The company describes the model as being designed for tasks that require “extended thinking,” positioning it as a specialized tool rather than a general-purpose chatbot. Architecturally, it is a Mixture-of-Experts (MoE) model, a design known for efficient scaling.
Uniquely, the model is not distributed in a standard format like SafeTensors. Instead, Moonshot has released the weights in a proprietary compressed format with a ‘.sbs’ extension. To use the model, developers must rely on the company's corresponding open-source inference framework, Kimi-Inference, which is designed to handle this specific format. The full model is available for download on Hugging Face.
Licensing and Limitations
The release is governed by the custom "Kimi-K2-Thinking License." While it permits academic research and general commercial use, it includes a significant restriction: the model cannot be used for any commercial purpose that directly competes with Moonshot AI's own products or services. This kind of non-compete clause places it in a category of controlled-use models, distinct from permissively licensed open-source software.
This release highlights a growing trend of companies making powerful models publicly available under specific, often restrictive, terms. For researchers, Kimi-K2-Thinking offers a new architecture to explore for advanced reasoning. However, the non-standard weight format and licensing terms present hurdles for broad commercial adoption and straightforward integration into the existing open-source ecosystem.
Sources
- Visit
moonshotai/Kimi-K2-Thinking
Hugging Face
More in Reasoning
Agnes-3.0-Flash arrives as a multimodal reasoning model
The new release pairs vision-language understanding with a hybrid-attention design aimed at long-context reasoning.

InternLM's Atria Dawn Preview Targets Agentic Tasks
A new mixture-of-experts model trained on verified tool interactions arrives as an early preview under an MIT license.
ZGCM-1 arrives as a fully open 7B reasoning model
A compact foundation model targets math reasoning and agentic search with tool use, and its makers are releasing it fully open.
0 comments
No comments yet. Be the first to weigh in.