Why open-source AI matters, and why private firms should support it
Open-source AI matters because it is the only kind you can run on your own hardware. For a law firm, a clinic, a defence supplier or anyone else whose data cannot go through someone else’s servers, open-weight models are the difference between using AI and not using it. We build on them, and we think every private firm has a reason to want them to keep existing.
What open means here
An open-weight model is one whose trained file has been published. Anyone can download it, usually from Hugging Face, the public hub where most of them are shared, and run it with open-source software such as llama.cpp. After the download, nothing connects you to the company that made it.
Strictly, open weights are not the same as open source: the file is public, but the data and code used to train it often are not. The practical freedom is the same. You hold the file and can run it without asking anyone.
Who cannot use the cloud
- Lawyers and notaries, bound by professional secrecy.
- Accountants and tax preparers, holding other people’s financial lives.
- Clinics and health professionals.
- Defence and security suppliers, whose contracts forbid sending data to outside services, and sometimes forbid an internet connection at all.
- Governments and public bodies that must keep data under their own law.
- Companies with trade secrets they would not email to a competitor’s supplier.
For these, a cloud AI with a strong contract is still someone else’s computer. An open model on their own machine is the only option that needs no trust in a third party.
What open models protect you from
- Changed terms. A cloud provider can change its price, its privacy terms or its rules at any time. A file on your drive does not change.
- A model that disappears. Cloud models are retired on the provider’s schedule. Work tuned to one model has to be redone. Your downloaded model answers the same way in five years.
- Foreign orders. Data held by a provider can be demanded from that provider. Data on your server can only be demanded from you.
- One supplier. If one company’s model stops suiting you, another open one can be loaded on the same machine in an hour.
Dependence on one company
On November 12, 2026, Jensen Huang resigned as chief executive of NVIDIA, the company whose graphics cards run most of the world’s AI, including most private servers. We do not know what that will mean for its products or its support for open software, and nobody else does yet either.
It is a reminder of how much of this field rests on a few companies and a few people. The answer is the same as for models: keep the pieces open and replaceable. The open-source runners already work on other makers’ hardware, so a private server is not tied to one card maker.
How to support it
- Use it. Every firm that runs an open model is proof that there is a market for them.
- Buy from companies that publish. Favour suppliers who release their models and tools.
- Share what you learn. We publish our measurements and methods on this site for that reason.
- Say so. When governments consult on AI rules, rules written only with cloud services in mind can make open models harder to release. Firms that depend on them should be heard.
Questions and answers
- What is the difference between open-source and open-weight AI?
- An open-weight model’s trained file is published and can be run by anyone. Fully open-source AI would also publish the training code and data. For running a model privately, open weights are what matters.
- Why does open-source AI matter for privacy?
- It is the only kind of AI that can run entirely on your own hardware, so it is the only kind that needs no trust in an outside provider.
- What is Hugging Face?
- The public website where most open-weight AI models are published and downloaded.