went through Cohere's final round twice, once in late 2025 (no offer, leveling mismatch) and once earlier this year (offer, accepted). sharing observations from both because I think the contrast is useful.
Format, 2026
4 rounds, all virtual, spread over one day or two half-days depending on interviewer availability. They try to group them but schedules slip. Round 1: Coding (2 medium problems, 60 min) Round 2: System design (LLM serving / infrastructure flavor, 60 min) Round 3: Behavioral (60 min, 4-5 questions, 1 interviewer + 1 shadow) Round 4: Cross-functional (product, research, or a mix, 45-60 min, more conversational)
What changed between my two loops
First loop (late 2025): the system design was more generic distributed systems, less LLM-specific. The behavioral round had more traditional competency questions.
Second loop (early 2026): the system design was clearly scoped around inference serving, context window management, retrieval-augmented generation at scale. The behavioral round had more of the values/judgment flavor others have mentioned (the 'what do you still second-guess' type).
My read: they updated the interview rubric somewhere between those two loops. The newer version is more specific to what they actually build.
Practical notes
Debrief and feedback loop: my recruiter gave me structured feedback after my first no-offer, which was genuinely unusual and helpful. They were specific about where I'd been weak (system design depth on inference specifically) rather than vague HR deflection. That feedback is what let me prepare better for round 2.
Post-onsite timeline: offer came 6 business days after the final round. One check-in from the recruiter at day 4 to say they were still in debrief.