Company interview guide
Z.ai (Zhipu AI) Machine Learning Interview Guide (2026)
Deep questions on GLM-style architecture and post-training, RL and GUI-agent design, and PyTorch implementation; very few reports so far.
Z.ai (Zhipu AI)'s machine learning interview loop has 4 reported stages (Architecture and post-training depth, Implementation round (PyTorch), RL infrastructure design, GUI-agent design and evaluation); typical length: Not reported.
Roles covered: Research Scientist, ML Engineer, Agentic AI Engineer and AI Engineer. The question bank holds 10 questions reported from Z.ai (Zhipu AI) interviews.
Try a free question from this loop's topics
Now Z.ai. Zhipu AI operates internationally under the Z.ai name.
- Interview process length
- Not reported
- Stages
- 4
- Reports linked to bank questions
- 10
- Bank report dates
- Sep 2026–Sep 2026
- Questions in bank
- 10 (0 free)
- Last reviewed
- October 2026
How the Z.ai (Zhipu AI) interview loop works
A public compiled write-up lists four technical rounds: depth on model architecture and post-training, an RL infrastructure design round, a GUI-agent design round and an implementation round in PyTorch. It does not give their order, length, format or language of the interviews, and we have no candidate report of recruiter, screening, behavioral or team-match steps. The company's careers page separates campus and experienced hiring and groups technical roles as algorithm and R&D, but describes no interview process. The outline below rests on that single compiled source, so confirm every round with your recruiter.
The Z.ai (Zhipu AI) interview process, stage by stage
01Architecture and post-training depth
Technical discussion; format and length not reported
A deep conversation on how large models like GLM are built and post-trained. The compiled questions ask about blank-infilling pretraining versus BERT and GPT, the trade-offs of a sparse mixture-of-experts with far fewer active than total parameters, multi-token prediction, bilingual Chinese and English data, and one model that serves both a thinking mode and a direct-answer mode.
What it tests
- Mixture-of-experts: why total and active parameters differ and what each costs in memory, compute and serving
- Multi-token prediction as a training signal and as a draft for faster decoding
- How GLM's autoregressive blank-infilling objective relates to encoder and decoder pretraining
- Building one model with thinking and direct-response modes, and the serving implications
- Tokenization, data mix and evaluation for a bilingual model
How to prepare
- Read the GLM-4.5 technical report and be ready to explain its main design choices in your own words.
- Work through MoE arithmetic by hand: parameters, active FLOPs per token and memory per device.
- For every technique, prepare both why it helps and what it costs at inference time.
No first-hand candidate report; one public compiled write-up (2026) lists this round with no format, length or order. The careers page names no interview steps.
02Implementation round (PyTorch)
Coding in PyTorch; tooling, length and whether it is live are not reported
The only coding round named in the compiled write-up. Its one listed question asks for a top-k mixture-of-experts router and then a comparison of auxiliary-loss load balancing with a loss-free, bias-based approach. Expect model components rather than general algorithm puzzles, though no report confirms that.
What it tests
- Writing a top-k gating function with correct softmax, selection and renormalization
- Computing per-expert load and an auxiliary balancing loss
- Explaining loss-free balancing by adjusting per-expert routing biases
- Clean, vectorized tensor code with correct shapes
How to prepare
- Practise writing small transformer and MoE components from memory with shapes annotated.
- After the code works, be ready to discuss capacity limits, dropped tokens and router collapse.
No first-hand candidate report; one public compiled write-up (2026) lists this round with no format, length or order. The careers page names no interview steps.
03RL infrastructure design
Design discussion; format and length not reported
A design round on the systems that run reinforcement learning for LLMs. The compiled write-up asks why long-horizon agentic RL calls for a disaggregated, asynchronous design, where rollouts and training run separately, rather than one colocated synchronous loop. Reward design also comes up, for example verifiable rewards versus a learned reward model.
What it tests
- Separating rollout generation from training and handling stale policies
- Throughput when agent rollouts vary widely in length
- Choosing between verifiable rewards and a learned reward model
- Failure modes such as reward hacking and stalled workers
How to prepare
- Be able to sketch an asynchronous RL pipeline: inference workers, a buffer, the trainer and weight syncing.
- Read up on slime, the asynchronous RL framework the compiled question names, and be ready to compare it with a synchronous setup.
No first-hand candidate report; one public compiled write-up (2026) lists this round with no format, length or order. The careers page names no interview steps.
04GUI-agent design and evaluation
Design discussion; format and length not reported
Design an agent that operates real apps from screenshots over dozens of steps, in the style of the company's AutoGLM agent and CogAgent model, or the cloud service behind it. The same write-up lists evaluation questions, on agents, agentic coding benchmarks and contamination, without tying them to a round; we place them here because an agent design needs an evaluation plan.
What it tests
- Perception, action space and state tracking for a screenshot-driven agent
- Error detection and recovery across a task of about 50 steps
- Evaluating an agent's trajectory rather than one response
- Reporting agentic coding and tool-use benchmark scores without fooling yourself
- Guarding against benchmark contamination
How to prepare
- Structure the design as perception, planning, action, verification and recovery, then add safety checks for irreversible actions.
- Bring an evaluation plan to every design answer: task success, step efficiency, variance across runs and contamination checks.
No first-hand candidate report; one public compiled write-up (2026) lists this round with no format, length or order. The careers page names no interview steps.
Questions and themes reported most often
- Explain MoE economics: total versus active parameters (public list) (architecture and post-training depth)
- Implement a top-k MoE router, then compare load-balancing methods (public list) (implementation round (pytorch))
- One model with thinking and direct-response modes (public list) (architecture and post-training depth)
- Multi-token prediction for training and faster inference (public list) (architecture and post-training depth)
- Why agentic RL needs an asynchronous, disaggregated design (public list) (rl infrastructure design)
- Design a screenshot-driven agent for a 50-step app task (public list) (gui-agent design and evaluation)
- Evaluate an agentic coding model on public benchmarks honestly (public list) (gui-agent design and evaluation)
What Z.ai (Zhipu AI) asks about
- LLM architecture & post-training35%
- Agent design & evaluation30%
- PyTorch implementation20%
- RL infrastructure15%
Z.ai (Zhipu AI) interview prep checklist
- Can explain the cost and benefit of a 355B-total, 32B-active MoE
- Coded a top-k MoE router with a balancing loss and explained a loss-free alternative
- Can describe multi-token prediction and how it speeds up decoding
- Sketched an asynchronous RL pipeline with separate rollout and training workers
- Designed a screenshot-driven GUI agent with recovery from failed steps
- Wrote an evaluation plan for an agentic coding model that checks for contamination
- Asked the recruiter about every round's type, length and language
Tips
- Ask your recruiter for the type, length and language of every round; public detail on this loop is very thin.
- Read the GLM-4.5 technical report and recent GLM releases; the compiled questions are built around them.
- Know mixture-of-experts arithmetic and routing well enough to both explain and code it.
- Prepare a clear account of hybrid reasoning: how one model learns when to think at length and when to answer directly.
- Be ready to design an asynchronous RL training system for long agent rollouts.
- Practise one end-to-end GUI-agent design, including error recovery and evaluation.
- Keep PyTorch fluency sharp: write model components from scratch without looking things up.
- Expect follow-up on anything on your resume; we have no report on behavioral rounds, so prepare a few project stories anyway.
Practice the Z.ai (Zhipu AI) loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Z.ai (Zhipu AI) interview FAQ
How many interview rounds does Z.ai (Zhipu AI) have for ML roles?
Candidates report 4 stages: Architecture and post-training depth, Implementation round (PyTorch), RL infrastructure design and GUI-agent design and evaluation. A public compiled write-up lists four technical rounds: depth on model architecture and post-training, an RL infrastructure design round, a GUI-agent design round and an implementation round in PyTorch. It does not give their order, length, format or language of the interviews, and we have no candidate report of recruiter, screening, behavioral or team-match steps. The company's careers page separates campus and experienced hiring and groups technical roles as algorithm and R&D, but describes no interview process. The outline below rests on that single compiled source, so confirm every round with your recruiter.
How long does the Z.ai (Zhipu AI) ML interview process take?
Reported timelines run Not reported.
What does Z.ai (Zhipu AI) ask in ML interviews?
By reported question mix: LLM architecture & post-training 35%; Agent design & evaluation 30%; PyTorch implementation 20%; RL infrastructure 15%. The most repeated themes: Explain MoE economics: total versus active parameters (public list); Implement a top-k MoE router, then compare load-balancing methods (public list); One model with thinking and direct-response modes (public list).
How should I prepare for the Z.ai (Zhipu AI) ML interview?
Ask your recruiter for the type, length and language of every round; public detail on this loop is very thin. Read the GLM-4.5 technical report and recent GLM releases; the compiled questions are built around them. Know mixture-of-experts arithmetic and routing well enough to both explain and code it.
Are there free Z.ai (Zhipu AI) interview questions?
LeetML has 10 questions reported from Z.ai (Zhipu AI) 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
- DeepSeekResearch- and systems-heavy loop: know the company's own papers, build model parts from scratch and reason about training at scale.
- Moonshot AIResearch-heavy rounds on long context, efficient attention, MoE and LLM serving, by the makers of Kimi; 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.
- BaiduResume-led technical rounds that mix ML fundamentals, a little math and classic data-structure coding.
- Mistral AIAttention coded from scratch in PyTorch, a heavy LLM fundamentals quiz and practical work against the Mistral API.
- ByteDanceSequential one-hour rounds that pair a deep resume grilling with a classic algorithm problem and ML fundamentals.