AI glossary · Beginner · Language models
Open weights
Open weights is the practice of releasing the internal numbers (weights) an AI model learned during training so that anyone can download and use them. You can run the model directly on your own computer or server.
About this term
- Level
- Beginner · Core ideas for first-time readers
- Category
- Infrastructure & using models
- Fields
- Language models
- Added
- 2026-10-07
How it works
An AI model consists of an architecture and weights (parameter values). An open-weight model releases these weight files so users can download them and run or fine-tune the model on their own hardware. By contrast, a closed model stays on the company's servers and can only be used through an API or app.
Why it matters
Open-weight models let organizations use AI internally without sending data outside, which is useful for companies and institutions where security matters. They are easy to adapt to one's own purposes and can also reduce costs in the long run. They also come up often in discussions of 'sovereign AI', which aims to run AI on a country's own technology.
Things to know
Open weights and 'open source' are not the same thing. Even when the weights are released, the training data or training code often is not, and licenses that restrict commercial use are sometimes attached. Also, running a large model yourself requires high-performance GPUs, so being released does not mean anyone can use it easily.
Closed models vs. open weights: what's the difference
| Closed model | Open-weight model | |
|---|---|---|
| Weight files | ✕Only on company servers | ✓Downloadable |
| How to use | Via API or app | ✓Run and fine-tune on your own hardware |
| Sending data outside | Sent to company servers | ✓Can be used internally only |
| Training data and code | ✕Not disclosed | △Often not disclosed |
| Caveats | Subject to company policy | △Licenses restricting commercial use; large models need high-performance GPUs |
Example
Mistral AI's 'Mistral Large 4', covered on this site, was first released as a preview, with the company saying the open weights would be released later. Like this, companies release models in various ways, such as launching a service via API first and releasing the weights afterward.
Where it came up on this site
- briefing Mistral AI releases 'Mistral Large 4' public preview; open weights planned for later this month 2026-10-07
- briefing Musubi releases 'PolicyLM-1.7B', an open-weight decision model for real-time content moderation (report) 2026-10-07
- briefing Reflection releases open-weight model 'Beam', aiming at an 'AI factory' for enterprises and nations 2026-10-06
- briefing T-Search: An open-weight multi-step search agent retriever based on Qwen3.6-35B-A3B 2026-10-06