Mistral AI · Primly Community

Went through the Mistral research eng loop last month, here's what actually happened

hardware_hugo · 4 replies

Applied for a research engineer role (Paris-remote hybrid). Here's the breakdown:

Round 1: Recruiter call, 25 min. Mostly logistics: what I'm looking for, visa status, timeline. She was direct and didn't waste time.

Round 2: Technical screen with a senior researcher. An hour. No LeetCode. We spent maybe 20 min on ML fundamentals: attention mechanisms, how MoE routing actually works, tradeoffs between dense and sparse models. Then 30 min on a practical problem about inference latency and how you'd approach reducing it in a production setting. I could tell they were evaluating whether I had opinions, not just knowledge.

Round 3: Coding + systems. Given a small Python task involving tokenization and called to walk through some inference optimization code I'd written in the past. They wanted to see how I reason about performance, not just whether I can code.

Round 4: Final with eng lead and one co-founder-adjacent person. Mostly a conversation. They asked what I found interesting about Mistral's technical direction vs. other labs. You have to have an actual answer to that, not a platitude.

Total loop: about 3 weeks. Got an offer. The process felt like talking to a research team that wants to hire collaborators, not just headcount. Definitely not a generic FAANG loop.

4 replies

newgrad_neil

did they ask you to read any of their papers beforehand, or was that just implied? trying to figure out how deep to go on the mixtral architecture before my screen

ml_mike

implied but definitely expected. i went through the Mixtral MoE paper and the original Mistral 7B paper before my screen. came up directly. read the abstracts and contributions at minimum, go deeper if you have the time.

corp_refugee

the co-founder-adjacent person in round 4 is interesting. at a team this size that's not unusual, but it does mean the bar is set at the top of the house. calibrate accordingly.

infra_ines

inference optimization in round 3 tracks with what i've heard. they're really focused on running models efficiently, not just training them. if you know quantization, speculative decoding, or batching strategies at a systems level that's worth surfacing early.