Background
Graduated from École Polytechnique and took a master's in AI at Carnegie Mellon, then became a research scientist at Meta's FAIR, where he was one of the leads on LLaMA — the release that made open-weight frontier models normal. He left with Arthur Mensch and Timothée Lacroix to found Mistral in April 2023; the company passed 100 employees by September 2024 and raised a multi-billion-euro Series D in September 2026 at a valuation above €20B.
He is the science half of a three-founder split (Mensch commercial/CEO, Lacroix engineering/CTO, Lample science). Publicly he is the least political of the three — his quotes are about model quality rather than sovereignty — but he fronts French AI events, including the opening keynote at the AI Now Summit in Paris in May 2026.
What they run now
Owns Mistral's research: the open-weight Mistral/Ministral families, Magistral reasoning models, Pixtral vision, Voxtral audio, and the post-training stack behind Le Chat. Inside his org, Mistral runs an in-house Human Data Annotation team (a Data Annotation Quality Specialist role sits in 'the Human Data Annotation team within the Science organisation') plus Technical Program Managers for Human Data Annotation whose job description is explicitly vendor-facing: manage contractor teams and vendor relationships across sourcing, contracts, onboarding, training and performance monitoring for data-collection campaigns with subject-matter experts, and build dashboards for throughput, budget and quality. The April 2026 Paris TPM posting is the clearest statement that Mistral buys externally supplied expert human data. Notably, Le Chat's image generation is outsourced to Black Forest Labs' FLUX, so Mistral's own visual-quality demand is smaller than its text and multimodal demand.
What they buy
Evidenced at org level rather than personally: Mistral runs data-collection campaigns with subject-matter experts, staffed through contractor teams and vendor relationships managed by TPMs in his organisation, with budget and quality tracked on dashboards. Modalities are text-first, with multimodal (vision, audio, document) growing; the Ministral 3 report describes curated multimodal SFT data and a pairwise reward model without naming a vendor. Design/taste data would be a new category for them rather than a refresh of an existing contract. Inference: as CSO he approves the programme and its budget but does not run individual campaigns.
Career
- 2023 – presentCo-founder & Chief Science Officer, Mistral AI — Heads the Science org, which contains the Human Data Annotation team.
- 2016 – 2023Research Scientist, Meta (Facebook AI Research) — Lead on the LLaMA open-weight model family.
- —Carnegie Mellon University, M.S., Artificial Intelligence
- —École Polytechnique, Engineering degree