In short: the weights are what a model has learned, stored as numbers. When a company publishes them, you can download them and run the model on your own hardware. That says nothing about what the license lets you do.
If you use a chat in your browser, you never hold the model. It sits with the provider, you send a question and get an answer. With an open-weight model the files are in front of you. You can run them on your own server, fine-tune them on your own data, or just look at how the machine is put together inside.
Weights and license are two different things
Downloading is one thing, the right to do whatever you like with what you downloaded is another. The fresh example is Kimi K3.1 Pro. On 6 November 2026 Moonshot AI published its weights on Hugging Face and, by its own account, calls it an open-weight model: 2.8 trillion parameters, 104 billion of them active for each token. Its license, though, is Moonshot's own, named kimi-k3.1-pro. The terms are not in the text we read, so we do not claim what they allow. All we know is that a license exists and that it is worth opening before you download.
By contrast, Anthropic's FlashProve code is under Apache 2.0, a license with well-known, widely read terms. There you do not guess. With a custom license you read.
Three questions for any license before you build on a model: can you use it in a paid product, can you fine-tune it and hand out the tuned version, and are there limits based on the size or kind of company. The answers are in the license text, not in the headline.
One more thing about scale. Downloading a huge model does not make it easy to run. 2.8 trillion parameters is serious-server territory, not laptop territory. So with the biggest models the question is rarely whether you can download them. More often it is whether you can afford to run them.
The visual is generated code-art, no third-party images.