What is an open-weight model?

An open-weight model is an AI model whose trained files, called weights, are published for anyone to download and run on their own computers. A company can run it privately, adapt it to its own data and build it into products, within the terms of its licence. The training data and training code are usually kept private, so it is not fully open source. DeepSeek, Qwen, Mistral and Meta’s Llama are well-known examples.

Why it matters for a business

Running an open-weight model means inputs never leave machines the company controls, no vendor can change the price or the model mid-deployment, and nothing updates until someone deploys it. In exchange, the company pays for capacity whether or not it is busy and owns the upgrades and the on-call rota. That capacity only undercuts per-call pricing at millions of requests a month, so most businesses that self-host are paying for the data path rather than a smaller bill. The licence matters as much as the download: some are ordinary permissive licences, others carry the lab’s own restrictions.

Go deeper

Open-source vs. closed AI models: what is actually open

On Atlas: Open-weight models, compared

Related terms

• Open-source AI
• Private LLM
• Fine-tuning
• Foundation model
• Inference

All terms, A to Z