Mistral AI CEO Arthur Mensch sat down with investor Elad Gil at Figma's offices to discuss how a four-person team built and shipped Mistral 7B in four months, starting from zero GPUs and scaling to 500 for training. Mensch, a former DeepMind staff research scientist, co-founded Mistral with ex-Meta researchers from the LLaMA project. The trigger was GPT: they saw a market gap, assembled a team, and executed a deliberate speedrun against better-resourced incumbents.
The core thesis driving Mistral's early product decisions is that small models were underserved. LLaMA 7B existed but underperformed. Mistral made it substantially better, targeting fast inference and low cost at high throughput. That decision drove immediate community adoption. The company then layered a commercial platform on top of open-source releases, including Mixtral 8x7B in December, and signed distribution deals with cloud providers including Microsoft Azure to reach enterprises that cannot use third-party SaaS due to compliance constraints. Financial services is the first vertical, chosen because it is the most procurement-mature.
The full conversation covers the structural tension between open and closed source AI, how Mistral organizes squads of five to ship research, what enterprise buyers actually need versus what the AI market assumes they need, and the state of the EU startup and regulatory environment. Mensch also previews vertical-specific open-source models and fine-tuning features coming to the platform. The operational details on team structure and the honest take on the Azure partnership make this worth watching in full.
[READ ORIGINAL →]