On July 16, Moonshot AI released Kimi K3. It’s a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window and native visual understanding. Moonshot describes it as the world’s first open model in the 3-trillion-parameter class, the largest open-weight release to date.
Most executives can’t say which AI models their teams use, or what data those models see. Kimi K3 makes that blind spot far more dangerous.
Until now, the open, downloadable models your teams could grab on their own were clearly weaker than the ones you pay for and control.
On the independent Artificial Analysis Intelligence Index, K3 is the strongest open-weight model published to date, within a few points of the closed frontier. As of July 28, 2026, it scores 57, fourth overall behind Claude Opus 5, Claude Fable 5, and GPT-5.6 Sol, and level with the tier just below them. Moonshot’s own evaluations put it ahead of several frontier models across coding and agentic benchmarks. The distance between open-weight models and the closed Western frontier is now measured in months.
And Moonshot has now published K3’s full open weights, putting a frontier-class model on infrastructure any capable enterprise can run itself.
New to some of these terms?
- Open-weight (open weights): A model whose trained values are published for download, so any organization can run it on its own hardware instead of only through the vendor’s service.
- Self-host: Running the model on infrastructure you control rather than calling a vendor’s cloud, so your data stays inside your own environment.
- Parameters: The internal values a model learns during training. More parameters generally means more capability, and the count is a rough proxy for a model’s size.
- Mixture-of-experts: A model design that splits work across many specialized sub-networks and activates only a few for any given task, so a very large model runs at a fraction of its full computing cost.
- Context window: The amount of text a model can hold in view at once, measured in tokens. A one-million-token window can take in thousands of pages in a single prompt.
- Token: The unit a model reads and writes in. A token is roughly three-quarters of a word, so 1,000 tokens is about 750 words.
- Benchmark: A standardized test used to compare AI models on a specific skill, such as writing code or answering technical questions.
- Agentic: AI that carries out multi-step tasks on its own, such as browsing, running tools, and acting on the results, rather than just answering a single question.
- Shadow AI: Employees using AI tools the organization hasn’t approved or can’t see, which puts sensitive data outside any policy or oversight.
- LLM gateway: A single controlled entry point that all AI requests pass through, so an organization can apply security, filtering, and logging in one place. LLM stands for large language model.