Company interview guide
DeepSeek Machine Learning Interview Guide (2026)
Research- and systems-heavy loop: know the company's own papers, build model parts from scratch and reason about training at scale.
DeepSeek's machine learning interview loop has 6 reported stages (Application and team routing, Paper and architecture deep dive, From-scratch implementation in PyTorch, Distributed training, low precision and serving, Algorithmic coding, Project deep dive and motivation); typical length: Not reported.
Roles covered: Research Scientist, Deep Learning Engineer, LLM Engineer and ML Engineer. The question bank holds 9 questions reported from DeepSeek interviews.
- Interview process length
- Not reported
- Stages
- 6
- Reports linked to bank questions
- 9
- Bank report dates
- Sep 2026–Sep 2026
- Questions in bank
- 9 (0 free)
- Last reviewed
- October 2026
How the DeepSeek interview loop works
A compiled public write-up lists four kinds of round for research and systems roles: a deep dive on DeepSeek's published architecture and training papers, from-scratch implementation in PyTorch, depth on distributed training and low-precision numerics, and an algorithmic coding round; it does not give the order, lengths or number of rounds. The company's careers site adds that teams are matched to candidates during the process and, for one infrastructure posting, that interviews favour a real project you built, reasoning through an unseen failure, and a conversation about AGI over memorised facts. No candidate has reported this loop to us, so the outline rests on very thin evidence.
The DeepSeek interview process, stage by stage
01Application and team routing
Online application for internship or full-time roles; any screening call is not described
Postings are grouped into broad families such as full-stack and algorithms, AI core systems, model data strategy and deep learning research. Several postings say you need not choose a sub-direction: the team is set from your background and wishes and may change as interviews go on, and one open posting routes resumes to whichever research or engineering team fits best.
What it tests
- Evidence of hands-on depth: shipped systems, reproduced papers, open-source commits
- Contest results (algorithm or HPC competitions) and academic record, listed as pluses in several postings
- Which team your background fits: training framework, kernels, data, agents or research
How to prepare
- Link a GitHub profile; one infrastructure posting says it looks at commit history rather than stars.
- State the directions you are open to, since placement is decided during the process.
- One kernel posting lists good Chinese communication as a core requirement, so confirm the interview language early.
The company's careers site describes direction-free applications, team placement during interviews and the profile it looks for; no candidate report describes this step.
02Paper and architecture deep dive
Technical discussion of DeepSeek's published models; length and format not reported
The compiled write-up stresses direct questions on the company's own papers: fine-grained mixture-of-experts with shared experts, load balancing without an auxiliary loss, multi-head latent attention, multi-token prediction, and the reinforcement-learning recipe behind its reasoning models, including group-relative advantages and why supervised fine-tuning was added back.
What it tests
- Explaining DeepSeekMoE versus a standard top-2 mixture of experts
- How latent attention shrinks the KV cache and what it costs
- Why multi-token prediction helps training and decoding
- GRPO versus PPO and what pure-RL training showed
- Keeping synthetic training data from causing model collapse
How to prepare
- Read the V2, V3 and R1 technical reports and be able to sketch each key idea on a whiteboard.
- For every technique, prepare the problem it solves, the trade-off and one way it could fail.
Only a compiled public write-up lists this round and its questions; no first-hand candidate report confirms it or its format.
03From-scratch implementation in PyTorch
Live implementation of model components in PyTorch; number and length of rounds not reported
Write core model pieces without library shortcuts. The one concrete prompt we hold is top-k mixture-of-experts routing with a shared expert, with follow-up on efficiency and correctness traps such as token dropping, capacity limits and load imbalance.
What it tests
- Top-k gating, dispatch and combine for a mixture-of-experts layer
- Spotting efficiency and correctness traps in routing code
- Clean, readable tensor code with correct shapes
How to prepare
- Practise writing attention, a mixture-of-experts layer and a training step from memory in plain PyTorch.
- Narrate shapes and complexity as you write, and test with tiny tensors.
Named in one compiled public write-up with a single example prompt; no first-hand candidate report describes it.
04Distributed training, low precision and serving
Discussion or design round on large-scale training and inference; format not reported
The compiled write-up lists depth on distributed training and low-precision numerics, with prompts on FP8 training at very large scale, overlapping computation with communication in pipeline training, and serving a 671B-parameter mixture-of-experts model within GPU memory. The careers site's training and inference framework postings name the same areas: model parallelism, long context, MoE, low-precision training and inference, RL systems and KV-cache strategies.
What it tests
- Expert, pipeline and tensor parallelism trade-offs for MoE models
- What breaks in FP8 training and how to keep it stable
- Hiding all-to-all communication behind compute
- Serving a huge MoE model under memory and latency limits
How to prepare
- Know the DualPipe and FP8 sections of the V3 report and the company's open-source infrastructure releases.
- Bring numbers: memory per parameter at each precision, KV-cache size and communication volume.
One compiled public write-up lists this round; the company's job postings name the same skills as duties but do not describe an interview stage.
05Algorithmic coding
Algorithm coding round; language, platform and length not reported
The compiled write-up includes one algorithmic coding round alongside the ML rounds. Several postings list algorithm-contest results as a plus, which suggests classic data-structures and algorithms problems, but we hold no example questions.
What it tests
- Standard data structures and algorithms under time pressure
- Correct, efficient code with edge cases handled
How to prepare
- Keep contest-style practice going: graphs, dynamic programming, heaps and binary search.
- Ask which language and environment the round uses.
Mentioned only in one compiled public write-up, with no example questions; no first-hand candidate report.
06Project deep dive and motivation
Conversation with follow-up questions; described on the careers site for one infrastructure role
One infrastructure posting on the careers site says interviews avoid memorisation and instead ask you to walk through something you really built, with probing follow-ups, to reason aloud through an unfamiliar failure scenario, and to share how you think about AGI. It is the only official description of interview content, and it is not specific to ML roles.
What it tests
- Depth and honesty about a project you personally built
- A clear reasoning path on a problem you have not seen
- Genuine, long-term interest in AGI research
How to prepare
- Pick one project and rehearse the hard follow-ups: what broke, why, and what you would change.
- Have a considered view on where model research is heading and why DeepSeek's open approach appeals to you.
Described on the company's careers site for one infrastructure posting, not ML-specific; no candidate report confirms it for research or ML roles.
Questions and themes reported most often
- Fine-grained MoE with shared experts versus a standard top-2 MoE (public lists) (paper and architecture deep dive)
- Multi-head latent attention and KV-cache compression (public lists) (paper and architecture deep dive)
- Multi-token prediction for training signal and faster decoding (public lists) (paper and architecture deep dive)
- GRPO, pure-RL reasoning training and why SFT came back (public lists) (paper and architecture deep dive)
- Implement top-k MoE routing with a shared expert in PyTorch (public lists) (from-scratch implementation in pytorch)
- Serve a 671B-parameter MoE model under GPU memory limits (public lists) (distributed training, low precision and serving)
- Synthetic training data without model collapse (public lists) (paper and architecture deep dive)
What DeepSeek asks about
- LLM architecture & DeepSeek papers40%
- Training & inference systems at scale25%
- From-scratch PyTorch implementation15%
- Algorithmic coding10%
- Project deep dive & motivation10%
DeepSeek interview prep checklist
- Can explain DeepSeekMoE, auxiliary-loss-free balancing, latent attention and multi-token prediction without notes
- Can contrast GRPO with PPO and summarise what the R1 work showed
- Implemented top-k MoE routing with a shared expert from scratch in PyTorch
- Can walk through FP8 training stability and compute-communication overlap at scale
- Sketched how to serve a very large MoE model within GPU memory and latency limits
- Prepared one project deep dive with honest answers on what failed
- Asked the recruiter for the number, type, length and language of every round
Tips
- Treat the company's technical reports (V2, V3, R1 and later) as the syllabus; most listed questions come straight from them.
- Be able to derive the memory and compute savings of latent attention and fine-grained MoE on paper.
- Write a mixture-of-experts layer with top-k routing and a shared expert in plain PyTorch until it takes minutes.
- Study FP8 training, compute and communication overlap and expert parallelism at the level of a framework engineer.
- Keep algorithm practice going; contest results are a stated plus and one round is reported as algorithmic.
- Make your GitHub activity and any reproduced papers easy to find from your resume.
- Expect team placement to be discussed during the process and say which directions you would accept.
- Confirm the interview language and the round list with your recruiter, since public detail is thin.
Practice the DeepSeek loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
DeepSeek interview FAQ
How many interview rounds does DeepSeek have for ML roles?
Candidates report 6 stages: Application and team routing, Paper and architecture deep dive, From-scratch implementation in PyTorch, Distributed training, low precision and serving, Algorithmic coding and Project deep dive and motivation. A compiled public write-up lists four kinds of round for research and systems roles: a deep dive on DeepSeek's published architecture and training papers, from-scratch implementation in PyTorch, depth on distributed training and low-precision numerics, and an algorithmic coding round; it does not give the order, lengths or number of rounds. The company's careers site adds that teams are matched to candidates during the process and, for one infrastructure posting, that interviews favour a real project you built, reasoning through an unseen failure, and a conversation about AGI over memorised facts. No candidate has reported this loop to us, so the outline rests on very thin evidence.
How long does the DeepSeek ML interview process take?
Reported timelines run Not reported.
What does DeepSeek ask in ML interviews?
By reported question mix: LLM architecture & DeepSeek papers 40%; Training & inference systems at scale 25%; From-scratch PyTorch implementation 15%; Algorithmic coding 10%; Project deep dive & motivation 10%. The most repeated themes: Fine-grained MoE with shared experts versus a standard top-2 MoE (public lists); Multi-head latent attention and KV-cache compression (public lists); Multi-token prediction for training signal and faster decoding (public lists).
How should I prepare for the DeepSeek ML interview?
Treat the company's technical reports (V2, V3, R1 and later) as the syllabus; most listed questions come straight from them. Be able to derive the memory and compute savings of latent attention and fine-grained MoE on paper. Write a mixture-of-experts layer with top-k routing and a shared expert in plain PyTorch until it takes minutes.
Are there free DeepSeek interview questions?
LeetML has 9 questions reported from DeepSeek interviews. The stage sections above link the free practice problems for the topics each round covers; the reported-question set itself is in the app.
Related interview guides
- Moonshot AIResearch-heavy rounds on long context, efficient attention, MoE and LLM serving, by the makers of Kimi; very few reports so far.
- Z.ai (Zhipu AI)Deep questions on GLM-style architecture and post-training, RL and GUI-agent design, and PyTorch implementation; very few reports so far.
- AlibabaResume-driven technical rounds that mix ML fundamentals, hand-written code and a hard look at your own projects, then HR.
- ByteDanceSequential one-hour rounds that pair a deep resume grilling with a classic algorithm problem and ML fundamentals.
- Mistral AIAttention coded from scratch in PyTorch, a heavy LLM fundamentals quiz and practical work against the Mistral API.
- OpenAIPractical coding, hands-on ML debugging and LLM system design, with a strong mission check.