Mistral AI CEO Arthur Mensch sat down with investor Elad Gil at Figma's office to cover the full arc of how Mistral was built: four co-founders from DeepMind and Meta, zero GPUs on day one, and a first model, Mistral 7B, shipped in four months on roughly 500 GPUs. The company's open core strategy, releasing Mistral 7B and Mixtral 8x7B publicly while selling commercial models and a managed platform, is the central tension worth understanding. Mensch is direct about why small models matter: inference speed, cost at scale, and a developer community that was already building on Llama 7B but needed something better.
The conversation gets specific where most AI founder interviews go vague. Mensch explains that his core research teams run in squads of four to five people, a structure he says mirrors every historically productive AI research group. He identifies two concrete go-to-market tracks: financial services enterprises accessed through cloud partnerships including a disclosed deal with Microsoft Azure, and developers at AI-native or AI-adopting companies reached directly through Mistral's platform. The Azure relationship matters because large enterprises cannot easily route data through third-party SaaS providers, and cloud marketplaces solve that compliance problem.
The roadmap section names specific near-term moves: new open source models targeting both general use and vertical-specific tasks, fine-tuning features on the platform, and an early-stage chat assistant called Le Chat that Mensch himself calls 'ChatGPT v0.' The EU regulatory angle and the broader argument for why a European AI lab is structurally necessary are in the full transcript and video, and they are worth the read for anyone tracking the open versus closed source debate beyond the US context.
[READ ORIGINAL →]