The labs behind AI models

Who builds AI models?

An AI model is the program that understands your request and writes the answer. Designing and training one takes specialized research teams, vast amounts of data and enormous computing power. That is why a small number of labs, often backed by big tech companies or major funding rounds, build the most advanced models.

They can be grouped by region:

• United States: OpenAI (the models behind ChatGPT), Anthropic (Claude), Google DeepMind (Gemini) and Meta (the Llama family).
• China: DeepSeek, known for strong open-weight models, and Alibaba with its Qwen model family.
• Europe: Mistral AI, a French lab offering both open and closed models.

Other players, more specialized or newer, appear in the table above. The country filter and the open or closed filter help you find your way.

Open model or closed model

A model’s weights are the countless settings it learned during training, in a sense its know-how. Two strategies coexist:

• Closed model: the weights stay secret. You can only use the model through the lab’s assistant or its API (the technical access point for developers). This is the case for the flagship models from OpenAI, Anthropic and Google.
• Open model (“open weights”): the weights can be downloaded. You, or a hosting provider of your choice, can run the model without going through its maker, adapt it to your needs and keep your data in-house. Meta, DeepSeek, Alibaba (Qwen) and Mistral AI release such models regularly.

Keep in mind that “open” does not always mean “free for any use”. Each license sets its own terms, sometimes restrictive for commercial use. Strategies also change: one lab may release some models openly and keep others closed. For a detailed selection, see our open source AI models page.

Why the lab’s country matters

Where a lab is based is not a minor detail, especially if you use its online service:

• Applicable law. Data protection rules, transparency or moderation obligations, and the ability of authorities to request access all depend on where the company is established and where its servers are. In the European Union, the GDPR and the EU AI Act govern these questions; other regions have their own frameworks.
• Your data. What you type into an assistant travels through the servers of the lab or its partners. Read its privacy policy to find out whether your conversations may be used to train future models.
• Content. A model partly reflects its maker’s choices and its home country’s framework: some topics may be handled differently, or avoided.

With an open-weight model running on your own servers, the data question looks different, since nothing is sent to the lab. This is general information: for sensitive business use, have your choices reviewed by a specialist.

What a lab profile contains

Each profile shows:

• the lab’s country and a short introduction;
• its list of models, with links to their pages in the AI models index;
• for each model, whether it is open or closed;
• where its models stand in the AI ranking, when they have been evaluated.

To compare models from different labs side by side, use the model comparison tool. And to see where this information comes from, read our methodology.