In April 2023, three researchers who had just left Meta and Google DeepMind founded a company in Paris with no product and no public model. Five months later, in September 2023, that company released Mistral 7B under the Apache 2.0 license: full weights, no usage restrictions, no request form to fill out. At seven billion parameters it was small enough to run on a single consumer GPU, and it outperformed larger models on most of the benchmarks anyone cared about at the time. That single release set the pattern Mistral has followed ever since. It was never going to out-spend OpenAI or match the training budgets coming out of Chinese labs. Instead it picked a fight it could actually win: licensing terms and language coverage.

A licensing bet made early

Mistral 7B mattered less for its size than for what came attached to it. Apache 2.0 is about as permissive as an open license gets: no restrictions on commercial use, no clawback clauses, no fine print about revenue thresholds that trigger a different set of terms. At a moment when several other labs were releasing weights under licenses with exactly those kinds of carve-outs, Mistral’s choice signaled that the company was competing for developer trust first and revenue second. That trust paid off. Mistral 7B became a default starting point for teams fine-tuning small models, not because it was the most capable option available, but because using it came with the fewest legal questions attached.

Eight experts, two active per token

In December 2023, Mistral released Mixtral 8x7B, the model that made mixture-of-experts architecture a mainstream open-source topic rather than a research curiosity. The idea is straightforward once stated plainly: instead of one dense network processing every token, Mixtral holds eight separate expert sub-networks, and a small router decides which two of the eight handle each token as it passes through. All eight experts sit in memory, but only two do the work at any given step, which means the model has the capacity of something much larger while keeping the compute cost of inference much closer to a smaller dense model. Four months later, in April 2024, Mistral pushed the same design further with Mixtral 8x22B, scaling the approach up to 141 billion total parameters while keeping the same two-active-experts-per-token routing. Both releases stayed on permissive licensing terms, which meant the broader open-source community could inspect, fine-tune, and redeploy a mixture-of-experts model at a time when most MoE work of comparable scale was locked inside proprietary APIs.

Coding agents and the sovereignty pitch

Later in 2024, Mistral released Devstral, a 24-billion-parameter model built specifically for coding agents that need to operate on private codebases rather than generic public repositories. It’s a smaller, more targeted release than the earlier headline models, and that’s the point: Mistral was no longer trying to win a generalist benchmark race. It was building a tool for a specific job, released under terms a company’s legal team could approve without a lengthy review, and paired with the same European emphasis on data handling and regulatory conformance that has run through the company’s positioning since its founding. Strong multilingual coverage, particularly across European languages, has been part of that same pitch: a model trained and governed under EU rules is a different purchase decision for a European enterprise than a model whose training data and legal jurisdiction are opaque or foreign.

None of this makes Mistral the largest lab in the field, and it isn’t trying to be. It’s closer to a regional bank competing against multinational banks: it can’t match their balance sheets, so it competes on terms of service and on being answerable to the jurisdiction its customers actually live in. That’s a real strategy, not a consolation prize. For more on why Apache 2.0 counts as genuinely open in a way many “open” licenses don’t, see Open Source vs Open Weight. And for the dense versus mixture-of-experts split explained in general, including why routing only a fraction of experts per token changes the economics of running a model, see Dense vs Mixture of Experts, which Mixtral’s eight-experts-two-active design put into practice well before the approach became common elsewhere.