Moonshot AI has released the weights for Kimi K3, a new model with 2.8 trillion total parameters. It can handle up to one million tokens at a time and work with both text and images.
The move gives developers more freedom to use the model outside Moonshot AI’s own products. Companies and research teams can download the weights, run the model on their own systems and adapt it for different jobs.
Kimi K3 is built to take on large amounts of information in one go. It can read long reports, contracts, research papers, project files and large codebases without constantly losing track of earlier details. That could make it useful in offices where the real challenge is not asking a quick question, but sorting through hundreds of pages and finding what matters.
It can also work with visual material. Users can upload charts, screenshots, scanned documents or software interfaces and ask the model to explain them, spot problems or connect what appears in an image with written instructions.
The 2.8 trillion figure sounds huge, but the full model is not active for every task. Kimi K3 uses a mixture-of-experts system, which calls on different parts of the model depending on the job. This cuts down the computing power needed during use.
Even with that setup, running the full version will require expensive hardware. Most people will not be installing it on a home computer. It is more likely to be used by cloud companies, large businesses and specialist developers, while the public accesses it through apps and online services.
The release matters for more than its headline numbers. Opening the weights gives developers a chance to test the model, improve it and build new products around it. It also reflects a wider change in the industry: the race is moving beyond chatbots and toward tools that can handle long, complicated work.
