This is a heads up to the definition of “Open source AI” by the Open Source Initiative. Thier position is that “Open Weights Are Good. Open Source Is Better.”

If you don’t get the difference yet, while preparing this post I noticed examples of people using “open source” when they mean “open weights”. You can check this video explaining what open source AI really means

For most of us the difference is “free” or “paid”… even if you have to spend resources to run open weight models. Well, the difference is larger and relevant to you: if something is truly open to be modified, then you can imagine the many variations that people will produce and publish, meaning that you can get your hands on a model that is better tailored to your needs. Therefore, even if you only care about using “free” models, you should care more about “open source” instead of only “open weights” models.

Examples of truly open source models include:

  • Olmo by Ai2, a Seattle based non-profit AI research institute founded in 2014. They develop foundational AI research and innovation to deliver real-world impact through large-scale open models, data, robotics, conservation, and beyond.
  • Apertus by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre, Switzerland’s first large-scale, fully open, multilingual language model.
  • Marin primarily developed by Open Athena, a nonprofit that empowers academic labs to build scientific foundation models.
  • SmolLM3 by Hugging Face, a platform where the machine learning community collaborates on models, datasets, and applications.