AI models come in two broad kinds. Closed models (such as Claude, GPT and Gemini) are only available through their maker's apps and API. Open-weight models (such as Llama, Gemma, Qwen, DeepSeek and many Mistral models) publish their trained weights, so anyone can download and run them. In our database, tick "Open weights" to see them.
Why choose an open-weight model
- Control over data. Running the model on your own servers means your data never leaves them. This matters in health, finance, law and the public sector.
- No lock-in. The model can't be retired or changed under you. You decide when to upgrade.
- Customisation. You can fine-tune the model on your own data and ship it inside your product.
- Low prices through APIs too. Because many companies can host the same open model, competition keeps API prices low. You don't have to self-host to benefit.
Why choose a closed model
- Top quality first. The most capable models on hard tasks (complex coding, long autonomous work) are usually closed, although the gap has narrowed a lot.
- Nothing to run. No GPUs, no servers, no updates to manage: you call an API and pay per use.
- Extra features. Built-in web search, file handling, safety systems, enterprise agreements and support.
The hidden cost of self-hosting
"Free to download" does not mean free to run. A large model needs expensive GPUs with lots of memory, and someone has to set up, secure, monitor and update it. For low or irregular traffic, paying per token through an API is almost always cheaper. Self-hosting starts to make sense with steady high volume, strict data requirements, or when a small model (which can run on a single GPU or even a laptop, especially with quantization) is good enough.
Check the licence
Open weights do not always mean open source, and licences differ. Some allow any commercial use; others restrict very large companies or certain uses. Read the licence on the model's Hugging Face page (linked from our model pages) before building a product on it.
A practical approach
Many teams use both: a closed flagship model for the hardest tasks and an open model, hosted or through a cheap API, for high-volume simple work. Prototype with whatever is fastest, measure quality on your own examples, then optimise cost.