What are open-weight AI models?

When an AI company releases an open-weight model, it publishes the model's trained parameters, known as weights, so anyone can download it and run it themselves. Closed models, by contrast, are only available through the company's app or API.

Open weights vs. open source

The two terms are often mixed up:

• Open weights: you get the finished model. You don't necessarily get the training data, the training code or the details of how it was made.
• Fully open-source AI: you get the model plus the data details and code needed to study and rebuild it. The Open Source Initiative published its Open Source AI Definition 1.0 on October 28, 2024, to spell out what qualifies.

Critics use the term "openwashing" for projects that call themselves open while sharing only part of the picture.

Well-known open-weight families

• Llama from Meta
• Mistral and Mixtral from Mistral AI
• DeepSeek V3 and R1, released under the permissive MIT license
• Qwen from Alibaba Cloud

Why people use them

• Run locally: no dependence on an outside company's servers
• Customize: fine-tune the model on your own data
• Privacy: sensitive data never leaves your machines
• Cost: often cheaper at scale than paying per use

The concerns

• Misuse: once weights are public, safeguards can be removed.
• Licensing: some "open" licenses restrict commercial use, so read the terms.