Mistral 3: A Heavyweight Model, Lightweight on Openness

Today, Mistral AI released the Mistral 3 family, including the massive 675B "Mistral Large 3." While this fills a gap for high-performance European models and aids local data governance, it fails to set a new standard for openness.

Mirroring its partner Microsoft (and OpenAI), Mistral has released model weights but kept training data under wraps. Without documented data, true performance evaluation and compliance testing are impossible. By conflating "open-weight" with "open-source," Mistral engages in "open-washing"—a practice that hurts the ecosystem by overshadowing projects that adhere to the EU-endorsed UN Open Source Principles.

Europe does not need more open-weight models; it needs a path towards data-at-scale to drive innovation. Mistral's approach sucks the oxygen out of the room for smaller, highly open alternatives that contribute to this mission like the Swiss AI Initiative's "Apertus" models, which offer the genuine openness and data access the industry requires.

This lack of commitment is jarring as Mistral AI recently signed an Open Source Initiative letter championing open data as a European value. By ignoring the very principles they publicly supported, Mistral has missed an opportunity to make a meaningful contribution to the European open source AI ecosystem. Europe needs AI-ready data, not more weights.