Mistral AI

Contribution to AI

Capability per parameter turned out to be a dimension worth competing on. The 7B model released in September 2023 — weights distributed by magnet link, under Apache 2.0, ahead of any paper — beat Meta's Llama 2 13B on every benchmark its authors ran at roughly half the size, pairing grouped-query attention with a sliding-window scheme that holds inference cost down over long sequences. Being small enough to fine-tune on a single rented GPU, and licensed without conditions or gating, made it the default substrate for the open community's derivative work; Zephyr, the OpenHermes series and thousands of other checkpoints were built on it within months. Mixtral 8x7B, three months later, did something no open release had managed: a sparse mixture of experts, forty-seven billion parameters resident but only about thirteen billion active per token, matching Llama 2 70B while serving several times faster. Routed sparsity stopped being a technique described in Google papers and became an artefact anyone could download, dissect and deploy. The second contribution is political rather than technical. A Paris laboratory arguing that obligations should attach to applications rather than to foundation models pushed the final EU AI Act toward exemptions for open releases, and turned European technological sovereignty from a speech into a company with weights on the internet.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).