“We are building the best AI in Europe, to make it available to everyone.”
The Story
Mistral AI was founded in April 2023 by Arthur Mensch (former DeepMind), Guillaume Lample, and Timothée Lacroix (both former Meta AI), in Paris. It raised €105 million in seed funding at a €240 million valuation — the largest European AI seed round — before shipping a single product.
In September 2023, they released Mistral 7B. A 7-billion parameter model that benchmarked above Llama 2 13B on every standard evaluation. The release came as a magnet link on X (formerly Twitter) in a post that simply said: "We are releasing our first model." No blog post. No paper. A magnet link. They were making a point about openness as both a principle and a product strategy.
Mixtral 8×7B followed — a sparse mixture-of-experts architecture that activates only 2 of 8 expert networks per token, giving inference efficiency of a 7B model while training quality of a much larger one. It outperformed GPT-3.5 on most benchmarks while costing a fraction of the compute.
Why They're in the Museum
Mistral's contribution to the museum's thesis is architectural: they proved that the dominant narrative ("bigger is always better") was a function of available resources, not evidence. The efficiency innovations in Mistral 7B and Mixtral — grouped query attention, sliding window attention, sparse expert routing — demonstrate that architectural intelligence can substitute for brute-force scale.
In the context of Katie's Law, Mistral represents constructive laziness applied to inference: don't compute tokens you don't need, don't activate experts that aren't relevant, don't ship a 70B model when a 7B model with better architecture achieves the same result. The laziness is designed into the system. That is the correct form.
The European Angle
Mistral also matters for regulatory and geopolitical reasons the museum cannot ignore: they represent the first credible European claim to AI frontier competitiveness. The EU AI Act, GDPR, and the question of AI sovereignty are not purely political — they are architectural constraints that will shape what AI systems can be deployed where, under what conditions. Mistral builds with those constraints as design parameters.
The most interesting thing Mistral did was release their first model as a magnet link. They were not just releasing weights. They were making a statement about what openness means.