Moonshot AI has released the full Kimi K3 model weights, opening its flagship AI model to downloads, private deployment, and fine-tuning. The release fulfills the company’s promise to publish the weights by July 27.
K3 is also appearing in third-party AI tools, but there is an important catch. Ollama currently offers it through the cloud, while LM Studio does not yet list an official K3 build. This is not a model most people can download and run on a MacBook.
The full model is about 1.56 TB
Kimi K3 has 2.8 trillion total parameters and activates 104 billion of them for each token. It supports text and images, has a 1-million-token context window, and uses a Mixture-of-Experts design that selects 16 of its 896 experts at a time.
Even with native MXFP4 compression, the official Hugging Face repository takes up roughly 1.56 TB. Moonshot’s deployment guidance recommends vLLM, SGLang, or TokenSpeed and a supernode with at least 64 accelerators.
That puts K3 in a completely different class from the smaller models people normally run locally on a Mac. Open weights make the model inspectable and self-hostable for well-equipped teams; they do not make it laptop-friendly.
Ollama access runs in the cloud
Ollama now lists kimi-k3:cloud, so users can launch the model from the familiar Ollama command line. However, the inference happens on Ollama’s servers rather than the user’s computer.
The listing currently requires an Ollama Pro or Max subscription and consumes extra usage credits. Ollama says it is adding capacity, which suggests access may remain constrained during the initial rollout.
LM Studio’s public model catalog does not show an official Kimi K3 build at the time of writing. Community conversions may appear, but K3’s size and new architecture mean finding a model in a catalog is not the same as being able to run it locally.
For comparison, models such as gpt-oss-20B can run offline on a suitably equipped Apple silicon Mac, while Gemma 4 includes models built for consumer hardware. K3 is aimed at data-center-scale deployment.
The license allows broad reuse, with conditions
The custom Kimi K3 License allows people and companies to use, modify, fine-tune, distribute, sublicense, and sell copies or derivatives of the model.
It adds extra conditions for large model-as-a-service businesses and commercial products with more than 100 million monthly users or over $20 million in monthly revenue. Those details matter more to AI providers than to researchers experimenting with the weights.
The release should make independent testing and private deployments possible for organizations with enough hardware. For everyone else, Ollama Cloud, Kimi’s own apps, and hosted APIs remain the practical ways to try K3.
