Company interview guide
Snap Machine Learning Interview Guide (2026)
Classic coding plus fast, broad ML fundamentals and recommendation-heavy ML design, with values questions in every round.
Snap's machine learning interview loop has 8 reported stages, from recruiter screen to values, project deep dives & q&a; typical length: 3-4 weeks.
Roles covered: ML Engineer, Applied Scientist, Deep Learning Engineer and Data Scientist. The question bank holds 48 questions reported from Snap interviews, 12 of them free to solve in your browser.
- Interview process length
- 3-4 weeks
- Stages
- 8
- Reports linked to bank questions
- 60
- Bank report dates
- Jan 2021–Sep 2026
- Questions in bank
- 48 (12 free)
- Last reviewed
- October 2026
How the Snap interview loop works
A recruiter screen (a few candidates met the hiring manager first), then a 60-minute technical screen that usually packs values questions, rapid ML fundamentals and an algorithm problem into one hour. Two 2026 candidates instead had two screen rounds, one ML design and one ML fundamentals, and some vision teams screen with an ML coding exercise. The virtual loop is four to six rounds of about an hour, in one day or split over two, drawn from general coding, AI-assisted coding, ML fundamentals, applied ML, ML system design, leadership, a product round at L6 and above and an informal Q&A. Some roles are tied to a team from the start and others match near the offer. Several candidates report slow recruiter replies or ghosting, so follow up politely.
The Snap interview process, stage by stage
01Recruiter screen
About 30 min call; a few candidates had a hiring-manager chat first
A conversation about your background, the project you are proudest of and why Snap. The recruiter also explains the loop and whether the role is tied to a team or matched later.
What it tests
- A clear summary of your most relevant ML project
- Why Snap and why this team
- Logistics: level, location, timeline
How to prepare
- Ask which round types your loop uses and whether any round allows an AI assistant.
- Ask whether there is an ML coding round; it appears in some teams' loops and not in others.
A recruiter or HR call opens the process in 10 ML-role and 8 data science reports; two ML candidates started with a hiring-manager chat instead. Snap's hiring page lists it as step one.
02Technical screen
Usually 60 min with one interviewer, sometimes with a shadow: about 10-20 min intro and values, 10-25 min rapid ML questions, 20-30 min of an algorithm problem
A dense hour that mixes three things. Reports describe quick-fire ML questions (overfitting by model family, imbalance, optimizers, metrics, attention) followed by a medium algorithm problem that can be pushed to hard; one candidate simply ran out of time on the coding. Some screens are coding-only or ML-only, and two 2026 candidates had two screen rounds instead, one ML design and one ML fundamentals with a resume deep dive.
What it tests
- Graph and grid problems: BFS shortest path, cycle detection, topological sort
- Expression parsing, string decoding DP and multi-list merging with heaps
- Short, precise answers on overfitting, bias-variance and class imbalance
- Explaining self-attention, multi-head attention and positional encoding
- Answering values questions concisely without eating the coding time
How to prepare
- Keep the ML answers to a minute or two each; the coding problem needs at least 20 minutes.
- Know cycle detection with three visit states and topological sort cold; one candidate failed on exactly that.
- Split overfitting answers by model family: linear models versus deep networks.
- Have Q, K, V shapes and why order information must be added ready to say out loud.
Free reported questions for this stage
- How Self-Attention Works, From Intuition to Tensor ShapesLLMs & GenAIMediumSolve
- Basic Calculator IICoding & AlgorithmsMediumSolve
- Median of Two Sorted ArraysCoding & AlgorithmsHardSolve
- CNNs Versus Transformers Under Data and Latency LimitsDeep LearningMediumSolve
26 ML-role reports describe the screen. 9 of the 15 since 2024 mixed values, ML questions and coding in one hour, 2 had two separate screen rounds, and the rest were coding-only, ML-only or project-focused.
03Virtual loop: coding
Usually 2 x 60 min, any language; Snap also lists an AI-assisted coding round, which no candidate report mentions yet
Algorithm and data-structure problems of medium to hard difficulty, often with a follow-up that changes the output or adds a constraint. Snap's own guide describes an AI-assisted variant where you solve a problem with an assistant and are judged on how deliberately you use it.
What it tests
- Search and backtracking, for example reaching 24 and then printing the expression
- Grid BFS with diagonal moves and edge cases
- Union-find connectivity and tree building from flat input
- Clean follow-ups that reuse your first solution
- Using an AI assistant to generate, debug and revise code when the round allows it
How to prepare
- Expect a follow-up on every problem; structure the first solution so it can be extended.
- Do not use AI tools unless the interviewer says the round permits them.
- In an AI-assisted round, narrate why you accept or reject each suggestion and test the result yourself.
Algorithm rounds in the loop in 15 ML-role reports, 8 since 2024, most often two rounds of medium to hard problems with follow-ups. The AI-assisted variant comes only from Snap's ML prep page.
04Virtual loop: ML fundamentals
60 min verbal round, rapid questions with follow-ups, often tied to your past projects
A breadth check across classical ML and deep learning, with persistent follow-ups. Reports mention cross-entropy and KL, dropout at train versus test time, why LLMs use layer norm, optimizer differences, regularization, calibration and imbalance, and candidates say neural network topics get special weight. Vision and generative teams have asked about GAN losses, StyleGAN, diffusion and 3D rotation math, and one 2026 loop added a discussion of recent papers.
What it tests
- Loss functions and their links: cross-entropy, KL divergence, label choices
- Normalization: batch norm versus layer norm, axes and inference behavior
- Optimizers: SGD, momentum, Adam and AdamW, and regularization trade-offs
- Calibration and class imbalance for a skewed classifier
- Transformer block structure and the reason for KL penalties in RL fine-tuning
How to prepare
- Answer each question with a definition, a consequence and a concrete fix; interviewers keep drilling.
- Be ready to defend modeling decisions from your own projects, which Snap names as a focus.
- Review deep learning basics broadly, not just your specialty.
Free reported questions for this stage
- Regularizers, Adam vs AdamW and the Decoder BlockDeep LearningHardSolve
- Logistic Regression: Gradient, Hessian, and ConvexityStatistics & InferenceMediumSolve
A breadth ML round in 17 ML-role reports, 9 since 2024, consistently fast-paced with follow-ups; vision teams add generative-model and math questions. Snap's ML prep page lists the round.
05Virtual loop: applied ML & ML system design
1-2 x 60 min discussion; Snap separates applied ML (model design) from system design (scale, latency, serving). In two 2026 reports it came as a screen round
Most prompts are recommendation or ranking problems on Snap-like surfaces: a lens carousel, a short-video feed with short-term interests, ads or a home feed. Other reported prompts include multimodal text-to-video search, harmful content detection and showing the top places in a moving map viewport. Interviewers probe labels, losses, cold start and the online versus offline split.
What it tests
- Framing the objective and choosing labels when engagement has several levels
- Candidate generation and ranking, with cold start for new items and users
- Embedding retrieval, contrastive training and its weaknesses
- Latency, freshness and when to use approximate nearest neighbours versus a model
- Monitoring, feedback loops and safety constraints
How to prepare
- Prepare one recommender blueprint you can adapt to lenses, stories, ads and short video.
- Be explicit about how weighted or multi-task labels turn into a loss and a final score.
- Know the limits of embedding-based retrieval and how you would mitigate popularity bias.
- Ask early whether the round leans toward modeling or infrastructure; one 2025 loop had one of each, and a 2023 candidate expected modeling trade-offs but got production recommenders.
Free reported questions for this stage
- Designing an AR Lens Recommender With Value-Weighted LabelsML System DesignHardSolve
- CLIP, Contrastive Losses and the Limits of Embedding RetrievalDeep LearningMediumSolve
Applied ML or ML system design in 15 ML-role reports, two of them as a screen round; 6 prompts were recommendation or ranking. Snap's ML prep page lists both round types.
06ML coding (team dependent)
60 min, Python or numpy, implement an ML building block from scratch; occasionally this is the screen itself
Not every loop has it. It shows up mostly for vision, generative and applied teams, in reports from the US, London, Vienna and Tel Aviv. The most reported task is writing a 2D convolution with padding, as plain loops or via matrix multiplication. Other reports mention blurring an image, a random crop, k-means and summing along an axis without library helpers.
What it tests
- Index arithmetic for convolution: padding, stride, output shape
- Vectorizing loops into matrix multiplication
- Image operations on raw arrays
- Explaining and debugging your output when the interviewer questions it
How to prepare
- Write conv2d both with loops and with an unfold-and-matmul approach until both feel routine.
- Compute and state output shapes before writing any loop.
ML coding from scratch in 12 ML-role reports from 2019-2026, 8 since 2024, at least one as the screen. It is not on Snap's published round list, so treat it as team dependent.
07Data science track: probability, stats, metrics
Data scientist loops: a screen with SQL and probability, then rounds on SQL, statistics, A/B testing and product metrics
Data scientist candidates describe a recruiter call, a screen with SQL and probability or a metric case, then four or five rounds on SQL, statistics, A/B testing, product sense, light Python, ML and behavioral, with few algorithm problems. The practice problems here come from an older public question list attributed to Snap, not from candidate reports, so treat them as representative rather than confirmed.
What it tests
- Conditional probability and counting on user panels
- Markov chains for active-user dynamics
- Likelihood ratio tests on lifetime data
- Diagnosing a drop in a key product metric
- Light Python coding on arrays
How to prepare
- Practise window-function SQL; candidates mention computing rates over a user's first sessions.
- Prepare an A/B test design for comparing two interfaces, including guardrail metrics.
Free reported questions for this stage
- Diagnosing a Drop in a Key Product MetricProduct & MetricsMediumSolve
12 data scientist reports, mostly 2021-2024: SQL plus probability in the screen, then SQL, stats, A/B testing and product rounds. The practice problems come from a compiled list, not these reports.
08Values, project deep dives & Q&A
About 15 min at the start of most technical rounds, project deep dives throughout, plus leadership or product rounds by level and a 30-min Q&A
Snap assesses values only through behavioral questions, and they appear in most rounds rather than in one dedicated interview; one candidate got an innovation question in every round. Expect at least one deep dive on a past project, sometimes framed as a retro on the ML system you know best, probing labels, loss, features, cold start and online versus offline design. Senior candidates may also get a leadership round and, at L6 and above, a product round, and some loops end with an informal Q&A with an engineer or TPM.
What it tests
- Stories that show Snap's values: kind, smart, creative
- Innovation: an idea you introduced, validated and drove to adoption
- A past project end to end: labels, loss, features, cold start, serving
- Ownership: what you personally did versus the team
- Handling ambiguity and cross-team influence
How to prepare
- Use Snap's Situation, Action, Impact, Learning structure and end each story with what you learned.
- Prepare several short stories so you do not repeat one across five rounds.
- Bring real questions for the Q&A round; it is not scored but it is your chance to judge the team.
Values questions inside technical rounds in 12 ML-role and 2 data science reports; 10 ML reports describe a project deep dive and 3 a Q&A round. Snap's hiring pages describe values questions and the Q&A.
Questions and themes reported most often
- Deep dive on a past ML project: loss, labels, features, cold start, online vs offline (10 reports, values, project deep dives & q&a)
- Rapid ML fundamentals: overfitting, imbalance, optimizers, metrics, bias-variance (8 reports, technical screen)
- Implement a 2D convolution or CNN layer from scratch, with padding (6 reports, ml coding (team dependent))
- Design a recommender or ranking system for a Snap-style surface (6 reports, virtual loop: applied ml & ml system design)
- Explain attention, multi-head attention and positional encoding (5 reports, technical screen)
- Grid BFS shortest path, including a closest meeting room variant (4 reports, technical screen)
- Generative models for vision teams: GAN losses, StyleGAN, diffusion (4 reports, virtual loop: ml fundamentals)
- Batch norm versus layer norm: axes, inference and why LLMs use layer norm (3 reports, virtual loop: ml fundamentals)
What Snap asks about
- General coding30%
- ML & DL fundamentals28%
- ML / recsys system design22%
- Behavioral & values9%
- ML coding from scratch7%
- Stats & probability (DS)4%
Snap interview prep checklist
- Solved course schedule, a grid BFS and a basic calculator variant in under 25 minutes each
- Can answer overfitting, imbalance and optimizer questions in two minutes each, with follow-ups
- Can explain batch norm versus layer norm with tensor axes and inference behavior
- Wrote a 2D convolution with padding and stride from scratch in numpy
- Sketched a lens or short-video recommender with cold start and multi-level engagement labels
- Prepared a project deep dive and several values stories, including one on innovation
Tips
- Treat the technical screen as three mini-interviews in one hour and practise that exact split against a timer.
- Drill medium-to-hard graph, BFS and parsing problems; they dominate reported coding questions.
- Build a one-page ML fundamentals sheet covering losses, normalization, optimizers, regularization, calibration and imbalance, and rehearse it aloud.
- Prepare a recommender design you can retarget to lenses, short video, ads or a feed, with explicit labels, losses and cold start.
- If your team does computer vision or generative work, expect ML coding and write conv2d from scratch until it is automatic.
- Know one of your own projects end to end: data, labels, loss, features, architecture, serving, monitoring and what you would change.
- Prepare five or six values stories, including one about innovation, since each round opens with behavioral questions.
- Confirm with your recruiter whether any round allows an AI assistant; otherwise assume it does not.
See the rest of the Snap loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Snap loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Snap interview FAQ
How many interview rounds does Snap have for ML roles?
Candidates report 8 stages: Recruiter screen, Technical screen, Virtual loop: coding, Virtual loop: ML fundamentals, Virtual loop: applied ML & ML system design, ML coding (team dependent), Data science track: probability, stats, metrics and Values, project deep dives & Q&A. A recruiter screen (a few candidates met the hiring manager first), then a 60-minute technical screen that usually packs values questions, rapid ML fundamentals and an algorithm problem into one hour. Two 2026 candidates instead had two screen rounds, one ML design and one ML fundamentals, and some vision teams screen with an ML coding exercise. The virtual loop is four to six rounds of about an hour, in one day or split over two, drawn from general coding, AI-assisted coding, ML fundamentals, applied ML, ML system design, leadership, a product round at L6 and above and an informal Q&A. Some roles are tied to a team from the start and others match near the offer. Several candidates report slow recruiter replies or ghosting, so follow up politely.
How long does the Snap ML interview process take?
Reported timelines run 3-4 weeks (About 3-4 weeks, often longer).
What does Snap ask in ML interviews?
By reported question mix: General coding 30%; ML & DL fundamentals 28%; ML / recsys system design 22%; Behavioral & values 9%; ML coding from scratch 7%; Stats & probability (DS) 4%. The most repeated themes: Deep dive on a past ML project: loss, labels, features, cold start, online vs offline; Rapid ML fundamentals: overfitting, imbalance, optimizers, metrics, bias-variance; Implement a 2D convolution or CNN layer from scratch, with padding.
How should I prepare for the Snap ML interview?
Treat the technical screen as three mini-interviews in one hour and practise that exact split against a timer. Drill medium-to-hard graph, BFS and parsing problems; they dominate reported coding questions. Build a one-page ML fundamentals sheet covering losses, normalization, optimizers, regularization, calibration and imbalance, and rehearse it aloud.
Are there free Snap interview questions?
Yes. LeetML has 48 questions reported from Snap interviews, and 12 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- MetaFast classic coding, an AI-assisted coding project and product ML design for recommendations and integrity.
- TikTokHour-long mixed rounds: a resume grilling, ML fundamentals and an algorithm problem, repeated two to four times.
- PinterestRapid-fire ML fundamentals, edge-case-heavy coding and embedding-centric recommendation design.
- GoogleClassic algorithm rounds plus ML design, ML domain depth and Googleyness, judged by a hiring committee.
- RedditA live model-building screen on a small dataset, then practical coding, ML fundamentals, ranking-system design and a hiring manager round.
- LinkedInProbability-flavoured coding, ML fundamentals and recommender design built around LinkedIn's own products.