How does the bootcamp + team match work in practice?

All Meta MLE / DS new hires complete a 4–6 week bootcamp covering Meta tooling (Manifold for production ML, FBLearner for training, Hive for data, pyTorch internals at Meta). During bootcamp you interview informally with teams. Strong performers get more team optionality — the top-ranked teams (Llama, Ads ranking core, Reels ranking core) typically take only the top bootcamp performers. Bootcamp performance is observable and is part of the team-match conversation.

Can I apply directly to Llama / FAIR?

Yes for senior+ research-engineer roles. FAIR and the GenAI org sometimes hire team-specifically rather than through the bootcamp + team match. The bar is research-engineer-shaped: published papers, open-source frontier-ML contribution, demonstrated research-engineering portfolio. Non-PhD candidates can clear this bar with substantial research-engineering work; new grads with no research portfolio typically cannot.

Is Meta DS more SQL-heavy than peer FAANG?

Yes. The Meta E3 / E4 DS interview has the highest SQL weight at FAANG. Meta runs on Hive (SQL-on-Hadoop) with internal extensions; SQL fluency at the level of medium-to-hard LeetCode-SQL problems with window functions and CTEs is non-negotiable. The Meta DS interview rubric explicitly tests SQL across multiple rounds; weak SQL at the screen stage is the most common screen-out.

How important is causal inference at Meta DS?

Required at E5+. Meta's product-DS work routinely uses propensity scoring, instrumental variables, difference-in-differences, and synthetic control. The canonical reference is Susan Athey's NBER work (athey.people.stanford.edu) and Hernán & Robins's 'Causal Inference: What If' (hsph.harvard.edu/miguel-hernan/causal-inference-book). Senior DS candidates who can't articulate the difference between an A/B test and a quasi-experimental design fail this dimension.

What's the on-call expectation for Meta MLE?

Significant at production-ML surfaces. MLEs on Feed / Reels / Ads ranking are in the on-call rotation for their model serving. The bar: identify whether a regression is a model issue or an infrastructure issue, run the rollback runbook, write the post-incident document. Research-engineer roles in FAIR have lighter on-call (typically pager only for training-infrastructure failures).

Is PyTorch fluency required at Meta MLE?

Implicitly assumed. PyTorch was originated at Meta (FAIR); production ML at Meta runs on PyTorch with custom internal extensions (FSDP at scale, custom Triton kernels, internal compile-mode features). Mid+ MLE candidates are expected to read PyTorch source to debug issues, write custom operators when needed, and know the framework's internals at a level beyond casual user.

Data Scientist / ML Engineer Hub

Data Scientist / ML Engineer at Meta (2026): Levels, Comp, Interview, Ranking and LLM Work

By Blake Crosley · Last verified 2026-04-29

In short

Meta is the largest production-ML company in 2026 by deployed-model count: Feed ranking, Reels ranking, Ads ranking, Marketplace ranking, Stories ranking, Llama foundation-model research, and dozens of internal ML platforms. Total comp at E3 (entry DS/MLE) clusters $200k–$290k, E5 (senior) $400k–$580k, E6 (staff) $620k–$900k, E7 (principal) $1.0M–$1.5M (levels.fyi 2026). Meta is unusual among FAANG for its bootcamp-and-team-match hiring model — you interview generically, then choose your team after offer. The Llama work (FAIR + GenAI org) is a research-engineer track distinct from production-ML.

Key takeaways

Meta E3 DS / MLE total comp $200k–$290k; E4 mid $280k–$400k; E5 senior $400k–$580k; E6 staff $620k–$900k; E7 principal $1.0M–$1.5M (levels.fyi/companies/facebook/salaries/data-scientist).
Meta uses bootcamp + team-match: candidates interview generically (no team-specific evaluation at offer time), do a 4–6 week bootcamp covering Meta-specific tooling, then choose a team. This gives stronger optionality than FAANG peers but reduces team-fit signal at offer.
Llama foundation-model work happens primarily in FAIR (research) and the GenAI org. Meta has open-sourced Llama 1 (2023) → Llama 4 (2025) → ongoing 2026 successor work, with the Llama 4 model cards published at huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct and adjacent.
Meta DS interviews are SQL-heavy and product-judgment-heavy. The E3 SQL screen is a canonical bar; Meta E5 product-DS interviews probe causal-inference depth (DiD, propensity scoring, instrumental variables) more than peer FAANG.
PyTorch was originated at Meta (FAIR) — staff-and-above PyTorch fluency is implicitly assumed. Real production-ML at Meta runs on PyTorch with custom-built distributed training (FSDP) and PyTorch 2 compile-mode features.

What DS and MLEs at Meta actually do

Meta's ML organization is the largest in production-ML by deployed model count and by infrastructure scale. The org is split across two distinct shapes:

Production-ML orgs. Feed ranking, Reels ranking, Ads ranking, Marketplace ranking, Stories, Threads, Instagram Feed, WhatsApp engagement. Each owns its own multi-objective ranker with its own training cadence, its own eval harness, and its own monitoring layer. MLEs here work on model-architecture iteration, feature engineering at scale (the offline feature store contains 10k+ features), eval design, and on-call rotation.
FAIR and GenAI org (research-engineer track). Llama foundation models, multimodal research, code models, agent research. This work is closer to AI-lab research-engineering — paper publication is part of the job. Llama 4 (2025) was a multi-team effort with public model cards at huggingface.co/meta-llama and a published research paper (ai.meta.com/research/publications). Meta is one of the few FAANG with publicly-open-source frontier foundation models.

The DS-vs-MLE distinction at Meta is sharp. DS at Meta is product-analytics — running A/B tests, defining metrics, partnering with PMs on product decisions. MLE at Meta is production-ML — training and shipping models. The career paths diverge at IC2/E3 and converge again at staff+. Meta DS is a great path for engineers who love SQL + causal inference + product judgment; Meta MLE is the path for engineers who love distributed training + eval design + production-ML systems.

The Meta interview and the bootcamp

Meta uses a generic-interview-then-team-match model that is unusual among FAANG. Process:

Recruiter call → 1 phone screen → 4–5 onsite rounds. No team affiliation at this point; interviews are generically scoped. Onsite for MLE: 2 coding rounds, 1 ML system design, 1 ML coding (Python + ML-specific), 1 behavioral. Onsite for DS: 1 SQL deep-dive, 1 product / case, 1 stats / experimentation, 1 behavioral.
Offer is generic. You receive a leveled offer (E3 / E4 / E5) but no team assignment. Compensation negotiation happens at this stage.
Bootcamp (4–6 weeks). All new hires go through a Meta-specific bootcamp covering internal tooling — Manifold (production-ML platform), the Hive query engine, FBLearner (ML training infrastructure), pyTorch internals at Meta. During bootcamp, you talk to teams, see what they're working on, and rank your preferences.
Team match. You're placed on a team based on team needs and your preferences. Strong performance in bootcamp gives more optionality.

What's distinctive: the bootcamp + team-match model gives candidates significantly more team optionality than at FAANG peers (where you interview into a specific team and the offer is contingent on team fit). The trade-off: no team-fit signal at offer time, so candidates can't negotiate based on team-specific demand.

Meta's interview prep references: Hello Interview's Meta-specific guide (hellointerview.com/blog/understanding-job-levels-at-faang-companies), the Meta engineering blog (engineering.fb.com), and the FAIR research publications (ai.meta.com/research/publications) for research-engineer prep.

Llama and the GenAI org: the research-engineer track

Meta's Llama work is the most public frontier-foundation-model effort outside the closed-AI-labs (Anthropic, OpenAI, Google DeepMind). Real production facts:

Llama family progression. Llama 1 (Feb 2023) → Llama 2 (Jul 2023) → Llama 3 (Apr 2024) → Llama 3.1 405B (Jul 2024) → Llama 4 series (Apr 2025, Scout 17B-16E and Maverick 17B-128E and Behemoth class) — all with public weights on Hugging Face Hub at huggingface.co/meta-llama and full research papers at ai.meta.com/research/publications.
FAIR (Fundamental AI Research). The research arm. Long-term horizon, paper-publication-driven, less product-aligned. Recent papers: 'Self-Rewarding Language Models' (Yuan et al., 2024), 'Toolformer' (Schick et al., 2023), 'LIMA: Less Is More for Alignment' (Zhou et al., 2023), and the ongoing Llama-team work.
GenAI org. Productizes the Llama work into Meta AI (the user-facing assistant), AI Studio (custom personas), and the AI surfaces inside Instagram / WhatsApp / Messenger.

Hiring on the Llama track is research-engineer-shaped — papers, evals, fundamental research engineering. Senior research engineers commonly have 5+ first-author or co-first-author papers at NeurIPS / ICML / ICLR. The bar for entry is closer to AI-labs (Anthropic, OpenAI) than to production-ML at Meta — non-PhD candidates need a substantive research-engineering portfolio (open-source frontier-ML contribution, co-authored paper, named GitHub project).

Compensation and negotiation at Meta

Meta compensation at all levels is heavy on RSUs (4-year vesting, sometimes 25/25/25/25 cliff, sometimes 10/20/30/40 backloaded depending on hire date). Total comp by level (per levels.fyi 2026):

Level	DS	MLE
E3 (entry)	$200k–$280k	$210k–$290k
E4 (mid)	$280k–$400k	$290k–$420k
E5 (senior)	$400k–$580k	$420k–$600k
E6 (staff)	$620k–$900k	$640k–$920k
E7 (principal)	$1.0M–$1.5M	$1.0M–$1.6M

Negotiation tactics that work at Meta in 2026: (1) competing offers from Google, Apple, or AI-labs are taken seriously and matched up to band caps; (2) sign-on bonus is the most-negotiable component; (3) RSU refresh is non-negotiable at offer but is significant at year 2+ if you're a strong performer; (4) at E5+, total-comp negotiation often pushes the candidate to the upper end of the band, but rarely outside the band. Stale negotiation tactics (asking for a higher level than offered) generally fail.

Frequently asked questions

How does the bootcamp + team match work in practice?: All Meta MLE / DS new hires complete a 4–6 week bootcamp covering Meta tooling (Manifold for production ML, FBLearner for training, Hive for data, pyTorch internals at Meta). During bootcamp you interview informally with teams. Strong performers get more team optionality — the top-ranked teams (Llama, Ads ranking core, Reels ranking core) typically take only the top bootcamp performers. Bootcamp performance is observable and is part of the team-match conversation.
Can I apply directly to Llama / FAIR?: Yes for senior+ research-engineer roles. FAIR and the GenAI org sometimes hire team-specifically rather than through the bootcamp + team match. The bar is research-engineer-shaped: published papers, open-source frontier-ML contribution, demonstrated research-engineering portfolio. Non-PhD candidates can clear this bar with substantial research-engineering work; new grads with no research portfolio typically cannot.
Is Meta DS more SQL-heavy than peer FAANG?: Yes. The Meta E3 / E4 DS interview has the highest SQL weight at FAANG. Meta runs on Hive (SQL-on-Hadoop) with internal extensions; SQL fluency at the level of medium-to-hard LeetCode-SQL problems with window functions and CTEs is non-negotiable. The Meta DS interview rubric explicitly tests SQL across multiple rounds; weak SQL at the screen stage is the most common screen-out.
How important is causal inference at Meta DS?: Required at E5+. Meta's product-DS work routinely uses propensity scoring, instrumental variables, difference-in-differences, and synthetic control. The canonical reference is Susan Athey's NBER work (athey.people.stanford.edu) and Hernán & Robins's 'Causal Inference: What If' (hsph.harvard.edu/miguel-hernan/causal-inference-book). Senior DS candidates who can't articulate the difference between an A/B test and a quasi-experimental design fail this dimension.
What's the on-call expectation for Meta MLE?: Significant at production-ML surfaces. MLEs on Feed / Reels / Ads ranking are in the on-call rotation for their model serving. The bar: identify whether a regression is a model issue or an infrastructure issue, run the rollback runbook, write the post-incident document. Research-engineer roles in FAIR have lighter on-call (typically pager only for training-infrastructure failures).
Is PyTorch fluency required at Meta MLE?: Implicitly assumed. PyTorch was originated at Meta (FAIR); production ML at Meta runs on PyTorch with custom internal extensions (FSDP at scale, custom Triton kernels, internal compile-mode features). Mid+ MLE candidates are expected to read PyTorch source to debug issues, write custom operators when needed, and know the framework's internals at a level beyond casual user.

Sources

About the author. Blake Crosley founded ResumeGeni and writes about data science, machine learning, hiring technology, and ATS optimization. More writing at blakecrosley.com.