Feb 29, 2024 - Technology

How Mistral became French for AI

Animated illustration of a robot putting a beret on its head and growing a goatee.

Illustration: Brendan Lynch/Axios

Paris-based Mistral — a nine-month-old startup with only a few dozen employees — is corralling enough investment and attention, including a high-profile Microsoft partnership, to put it in the top tier of AI companies globally.

Why it matters: Mistral is much younger than OpenAI, yet the French company's top model already rivals GPT-4's performance in accuracy and common sense reasoning.

Driving the news: In the past week, Mistral launched its Mistral Large model, which now ranks second only to GPT-4 in several industry benchmarks.

  • IBM announced Thursday it will offer an optimized version of the Mixtral-8x7b model on its WatsonX platform that increases throughput (the amount of data that can be processed in a given time period) by 50%, and which cuts latency by 35-75%.
  • In addition to the Microsoft deal, under which Mistral made its models available to Azure cloud customers while giving the giant a small stake, Mistral also has a deal with Amazon, making it the seventh foundation model developer available to Amazon Bedrock customers.
  • Mistral also launched Le Chat, a "multilingual by design" chatbot.

Catch-up quick: Mistral's funders include Nvidia and Salesforce, in rounds led by Andreessen Horowitz that value the company at $2 billion.

  • The company's founders are led by 31-year-old CEO Arthur Mensch, a former AI engineer at Google DeepMind, and Timothée Lacroix and Guillaume Lample, former classmates of Mensch's who worked at Meta's Paris AI lab.

The big picture: The stars are aligning for Mistral for several reasons.

  • It suits big companies such as Microsoft to bet on a French startup: They're looking to reassure regulators around the world of their good faith, and they also want to tap different language markets.
  • Some of Mistral's models are open, providing a counterweight to Microsoft's bets on closed models such as OpenAI.
  • Alongside Google's Gemini model stumbles, Mistral is basking in its models' reputation for producing relatively unopinionated responses.

Context: The French tech scene has been gaining scale and momentum in recent years, propelled in part by the overt embrace by President Emmanuel Macron, who has sought to establish Paris as the EU's tech hub after Brexit.

  • HuggingFace, founded by three French engineers, has become the AI industry's go-to scoreboard and communal space.

What they're saying: Mistral speaks the language of Silicon Valley VCs.

  • Mensch says he offers "unparalleled value" to customers and told the Wall Street Journal his ambition is to create "the most capital-efficient company" in AI.
  • Mistral Large cost only around $20 million to train, Mensch told the Journal — compared to $50 million or more for OpenAI's models.

Reality check: Microsoft's investment in OpenAI is around 800 times greater than its investment in Mistral: $13 billion versus $16 million.

The intrigue: Mistral was lobbying in both Paris and Brussels just weeks after it launched, as officials negotiated the bloc's comprehensive AI Act.

  • The firm's unusual lobbying effort helped to water down requirements on developers of general-purpose AI, mirroring goals of some of the big tech companies Mistral now partners with,
  • Eyebrows are raised in Brussels, as some officials and journalists suggest Mistral may have been acting as a big tech lobbying front.
  • The tables turned when Mistral announced its Microsoft deal: the EU will look into the agreement for any evidence it's designed to limit competition.

One fun thing: Mistral is good at picking names.

  • "Le Chat" is a play on words — the chatbot's name is also French for "cat."
  • The company's name means "masterful" in a southern French dialect — and also refers to the northwesterly wind that whips through southern France.
Go deeper