Definition
Open-weight model is an open-weight model is an AI model whose trained parameters (weights) are published so anyone can download and run it, though the training data and code may stay private.
A model's weights are the numbers it learned during training. They are the model. If a lab publishes them, you can download the model and run it on your own hardware or a cloud server you control, instead of sending every request to the lab's API.
That is an open-weight model. Well-known examples come from several labs, and each ships under its own license.
The two phrases get mixed up, but they are not the same. Traditional open-source software gives you the source code and the right to change and redistribute it. With AI models, "open weight" usually means you get the finished weights and the right to use them under a license, but not necessarily the training data, the training code, or the full recipe.
Licenses vary a lot. Some are permissive and allow almost any use. Others add limits, such as restrictions on very large companies or on certain uses. Always read the license before building a product on a model.
Running a model yourself means you handle the hardware, the serving software, updates and safety filtering. Big models need a lot of memory, especially a mixture of experts design, where every expert must be loaded even though few run at once. For many teams, a hosted API is still simpler.
Open weights also cannot be taken back once released. That is a feature for users and a concern for some safety researchers.
Open-weight models are used for inference on your own machines, and each has its own context window limit, set by how it was built.
Choose open weights when you need control over data, cost or customization. Choose a hosted model when you want the least work.
No. Open weights means the trained model is downloadable under a license. Open source also implies access to code and often data, which open-weight releases may not include.
It depends on the license. Some allow almost any commercial use, others add conditions. Read the license first.
Mistral opened a public preview of Mistral Large 4 on Tuesday, a mixture-of-experts model with about 1 trillion parameters and 49 billion active, and promised weights by the end of October. The API is a preview. The weights are the launch.
Reflection AI introduced Beam on Monday, a 501-billion-parameter mixture-of-experts model with 23 billion active, and promised Apache 2.0 weights later in October. Until the weights ship, it is a waitlist and a claim about efficiency.