Mistral Launches Mixtral 8x7B, Open AI to Rival GPT-3.5

Mistral AI released Mixtral 8x7B today, an open large language model that, according to early results published by the company itself, matches or outperforms OpenAI's GPT-3.5 and Meta's Llama 2 70B across most standard industry benchmarks. The French startup, founded just seven months ago by former Meta and Google DeepMind researchers, timed the announcement to coincide with the close of a funding round worth roughly €385 million.

What a mixture of experts is

Mixtral 8x7B isn't a monolithic model — it's what's known as a "mixture of experts" (MoE). Rather than firing up the entire neural network to process every word, the system splits the work across eight specialized sub-networks of 7 billion parameters each and routes every token to just two of those eight via a routing mechanism. The result is a model with a total capacity in the tens of billions of parameters that, in practice, only taps a fraction of that capacity for any given calculation — dramatically cutting compute costs compared with a dense model of equivalent size.

The architecture isn't Mistral's invention: Google has been exploring it in research for years, and there's been speculation for months that models like GPT-4 use some variant of mixture-of-experts internally, though OpenAI has never confirmed it. What sets Mixtral apart is that Mistral chose to release it openly, with downloadable weights — something neither OpenAI nor Google does with their flagship models.

A startup betting on openness

Mistral AI was founded in Paris this past May by Arthur Mensch, Guillaume Lample and Timothée Lacroix, three researchers who previously worked at Google DeepMind and Meta. Since its first release — the Mistral 7B model launched in September — the company has championed an open-weights strategy that sets it apart from the major U.S. labs: while OpenAI, Anthropic and Google keep their most powerful models closed off and accessible only via API, Mistral releases its models for anyone to download, modify and run on their own infrastructure.

That bet has turned the French startup into Europe's reference point in the debate over open versus closed AI, just as the European Union hammers out the final details of its AI Act — a law in which general-purpose models have become one of the most contentious issues.

A round that lives up to expectations

Alongside the model launch, Mistral confirmed it has closed a funding round worth roughly €385 million — a notable sum for a company barely seven months old that had already raised €105 million in a seed round back in June, led by Lightspeed Venture Partners, at the time the largest seed round ever recorded in Europe. The fresh capital puts Mistral among Europe's best-funded AI startups and reinforces its stated ambition to compete directly with OpenAI, Anthropic and Google on foundational models — while pursuing a business model that pairs open model releases with paid commercial services for enterprises.

Why it matters

Until now, anyone wanting an open model with performance close to GPT-3.5 had to turn to Meta's Llama 2 70B, which demands far more memory and compute power to run. If Mistral's figures hold up under independent scrutiny, Mixtral 8x7B would deliver similar or better performance at a notably lower inference cost, since it only activates part of its parameters at each step. That makes it cheaper for developers and companies to run the model on their own servers instead of relying on an external API.

The launch also marks a symbolic milestone: it's the first time an open model from a European startup has placed itself, by its own account, on par with the flagship OpenAI product that popularized ChatGPT. It remains to be seen how the independent community evaluates these comparisons, and what practical limitations emerge once outside developers start deploying the model beyond Mistral's own lab.