just finished the Mistral AI onsite a few weeks ago. took the offer so I can share freely now.
the final round was fully remote, four back-to-back hours with a 15-minute break in the middle. here's each segment:
round 1: system design (60 min) covers AI/ML serving infrastructure. I got a prompt about scaling a real-time inference endpoint. see other posts here for what to prep, but know it's not generic distributed systems. vocabulary around GPU batching, token throughput, latency percentiles. the interviewer will push you on cost and failure scenarios.
round 2: take-home debrief (45 min) this is a live walk-through of the code you submitted. they had clearly read it. they started with two questions about decisions I'd made, then we did some live coding in it, extending it slightly. be able to talk about every choice. also be ready to say 'I'd do this differently now' without being defensive. they responded well to that.
round 3: second coding problem (45 min) more open-ended than the phone screen problem. started from a vague prompt, I had to scope it before writing anything. that scoping conversation was probably half the signal. they weren't looking for the most optimal implementation, they were looking for whether I could define a solvable problem from a fuzzy one.
round 4: behavioral / values (45 min) covered in another post here. substantive questions about how I handle ambiguity, disagreement, and technical responsibility.
debrief timeline: I heard back in 5 business days. offer call came from the hiring manager directly, not the recruiter. that was a nice touch.
overall: the hardest part of the day wasn't any single problem, it was maintaining coherent reasoning for four hours straight. they don't seem to be looking for someone who crushes every round, more for someone who thinks well consistently. felt like a genuine evaluation rather than an obstacle course.