Open-weight AI models: what open weights means and why it matters
Open-weight models are AI models whose parameters (weights) the developer releases publicly. You can download them, run them on your own device and fine-tune them, unlike closed models that are only available through a service.

In plain words
A closed model is like a restaurant: tasty, but they will not give you the recipe and you can only eat there. An open model is like a dish they pack for you to take away: you can heat it at home, add your own spices, share it with friends. A fully open-source model also comes with the recipe and the list of ingredients showing how it was made.
How it works
- The developer trains the model and puts the parameter file online.
- The licence defines what you can do with it: for study, for business, with restrictions.
- Users download the model and run it on their own computer or server.
- The community makes compressed, fine-tuned and improved versions.
- The data the model was trained on usually stays closed.

Where you'll see it
Well-known open models include Llama, Mistral, Qwen, DeepSeek, Gemma and others. People run them with local AI apps on laptops, and they are used in startups and on smartphones. Some AI features on phones also run on small open models.
Does it matter when buying
Yes, if privacy matters to you: a local model does not send your data to a server. For this you need a computer with enough memory. But keep in mind that open models are usually weaker than the best closed ones, and you are responsible for how you use them.
Myths
- Myth: An open model is the same as open source. Reality: Not always. Often only the weights are open, not the training code or data.
- Myth: You can use open models however you like. Reality: Read the licence: some forbid commercial use or certain scenarios.
FAQ
Can I run an open model on a phone?
Small ones, yes, through special apps, if the phone has enough memory.
Are open models safe?
The model itself is a file of numbers. Download it from official sources to avoid fakes.
Are they free?
Downloading is usually free, but you need your own hardware or a rented server.
Related terms
Model parameters, Quantisation, Fine-tuning, On-device AI, Large Language Model