Definition
An AI model whose weights are published and downloadable, allowing it to run on your own infrastructure without depending on its creator's cloud.
An open-weights model (Llama, Mistral, Qwen, DeepSeek, GPT-OSS…) can be downloaded and run anywhere: on a company server, a private cloud, even a workstation. The processed data never leaves the machine, unlike proprietary models accessible only through a cloud API.
The best open models now rival proprietary ones on the majority of business tasks, at a controlled operating cost. They are the keystone of sovereign deployments: even if every American provider cut off their APIs tomorrow, a local model would keep working.
An SME downloads a model such as Llama or Mistral and runs it on its own server. No request goes to the model's publisher, and nobody can cut off access or change the terms overnight.
When you want to control your data, your costs and your dependence on a vendor, or to work without a connection.
Is an open-source model free?
The model itself usually is, depending on its licence. What remains is the hardware to run it, the installation and the maintenance. In use, there is no bill per user or per request.
All AI models, deployed on your infrastructure, with your documents connected.