Moonshot AI released a new AI model called Kimi K2.5 in January 2026 that anyone — researchers, developers, companies — can download and run on their own computers, rather than relying on a company's servers, according to Moonshot AI's GitHub repository.

The model contains 1 trillion parameters, which you can think of as connections or "memory units." But here's the important part: it only uses 32 billion of those during any single task. This design, called a mixture-of-experts, is like having a team of specialists in different areas. When the model needs to do something, it activates only the specialists relevant to that task, rather than waking up everyone at once. This keeps costs and processing time lower than models that use all their parameters every time.

The new model can also understand images, not just text. It can handle very long documents — up to 256,000 tokens (think of tokens as rough word units) — compared to 128,000 as a rough industry standard. It can also call external tools and software automatically. These features are designed to let multiple AI instances work together on complex problems, per Moonshot AI's API documentation and HPCwire's coverage.

What Makes This Different

This model is built for companies that want to run AI on their own computers or private networks, rather than sending their data to someone else's servers. That matters when a company has confidential information they cannot send elsewhere. Open-weight models let teams do that.

The model has been specifically trained to help with coding tasks, which is valuable because code is easy to check for correctness — you can run it and see if it works. Many companies are using AI to automate routine coding work, and this model targets that use case directly.

How This Model Can Be Used

When weights are open, it means people can download the model's underlying parameters and run it themselves. It does not necessarily mean all the training code or exact data is public. But download access alone opens up real possibilities: teams can fine-tune the model for their specific needs, test it in detail, and deploy it on their own hardware.

Moonshot AI is a company based in Beijing that has already built Kimi as a product for users in China. This open-weight release brings that work to developers and companies around the world. The real question now is whether all these features work well when deployed at large scale in the real world — something the broader developer community will test much faster than any lab test can.