Machine Learning Internships in 2026: Three Very Different Jobs, One Title

Read this before you apply to fifty of these

Most pages ranking for “machine learning internships” are job boards counting listings — “2,000+ open now” — or guides that tell you to learn Python and apply broadly. Neither will tell you the thing that actually decides your summer: the title “ML intern” is stapled onto three different jobs with different pay, different screens, and different odds. Apply to a research internship with a Kaggle-only profile and you’ll lose to someone with a professor’s letter. Apply to a quant firm with a solid GPA and no competition math and you’re buying a lottery ticket priced in weeks of your life.

So this page sorts the field the way employers actually sort it, tells you what each track has interns do all day, what the application technically requires, and what to do — specifically as an ML student — if you come up empty. For the full dated application calendar across every employer, use the AI internships hub; this page reuses only the dates that change your next move.

The three tracks (and which one you’re actually eligible for)

Track 1 — Research internships: the pipeline into ML as a discipline

This is the track people picture when they say “ML internship”: you’re attached to a research group, you read papers, you run experiments and ablations, you help push something toward a workshop paper or an internal report. The day-to-day is slower and more open-ended than product work — a week can be “we tried the obvious thing and it didn’t work, here’s why” — and that’s the job, not a failure of it.

The real options, as of July 2026:

• NSF REU sites (ML/AI) — the structured undergrad research route. Roughly $6,300 for a 9–10 week summer, plus housing and travel at most sites. US citizens and permanent residents only. Applications run November–March, with most deadlines January–March 2027 for summer 2027 — this is the one track where the clock starts later than big tech’s, so it doubles as a second chance if the fall cycle goes badly.
• Google DeepMind Student Researcher — the realistic “big lab research” door for students. Open to enrolled BS, MS, and PhD students in CS or related fields, 12–24 weeks at 4+ days a week, in person. Paid, though Google doesn’t publish the rate on the program page. Rolling — watch it from fall 2026.
• Meta FAIR / GenAI research internships — PhD students only. If you’re an undergrad or a master’s student, this door is closed; don’t spend an application on it. Meta’s general SWE internship (postings go live around early September 2026) is the version you can actually get.
• Ai2 (Allen Institute for AI) — paid research internships for undergrads and grad students, ~12 weeks on-site in Seattle, year-round and rolling. Less famous than the big labs, which is exactly why it’s worth your application.
• MLH Fellowship — remote, 12 weeks, beginner-friendly, mentored open-source work with an educational stipend and need-based living support up to $5,000. Not a research internship in the strict sense, but it reads as structured experience on a resume when you have none.

What this track screens for: evidence you can survive ambiguity — a course project that went beyond the assignment, any research assistant time, a professor who’ll vouch for you. Grades matter more here than anywhere else on this page, because a transcript is the only standardized signal research groups have.

Track 2 — Applied and product ML at big tech: the volume track

This is where most ML internships actually live. You’re on a product team, and the work is honest engineering: building and cleaning data pipelines, writing evaluation harnesses, fine-tuning and benchmarking models someone else designed, and shipping the boring 80% that makes a model usable. Interns who expect research and get data cleaning are a cliché in every intern cohort; go in knowing it and you’ll do well.

The employers running ML-relevant intern programs for the 2027 cycle: Apple’s AIML undergrad internship (CS, CE, data science, applied math majors; applications September–November 2026, rolling by team), NVIDIA (roles open August–October 2026, rolling by lab — and NVIDIA’s own page says it weighs practical skills, projects, and research work alongside GPA), Amazon’s Applied Scientist and SDE intern roles (the earliest big opener, July–August 2026), Microsoft (opens around mid-August 2026), and Google — STEP for first- and second-years plus the Student Researcher program above, with Google’s brief 2–4 week application window landing around mid-October 2026.

Pay clusters around $45–62/hour for big-tech ML interns — treat that as an estimate band, not a quote; companies publish “paid and competitive” and the numbers come from aggregators. That’s roughly $8k–11k a month, plus relocation or housing support at most of these programs.

What this track screens for: software engineering fundamentals first, ML second. You’ll pass or fail on coding rounds and one real project you can defend — more on the bar below.

Track 3 — Quant ML: the documented pay ceiling, priced accordingly

Quant trading firms hire ML interns, and they are the highest-paid internships that exist — not just in ML, in anything:

• Jane Street — roughly $4,500–6,000/week (annualized around $300k), plus something like $15–20k of in-kind housing and flights. Opens August 2026, rolling.
• Citadel / Citadel Securities — roughly $4,300–5,800/week (~$47–64k for an 11-week summer), plus a $15–25k signing bonus and housing. Opens around August 2026, rolling.
• Two Sigma — roughly $19,500–28,000/month for quant/ML interns, and the intake skews toward grad students and PhDs.

Now the number that matters more than the pay: Jane Street takes roughly 250 interns from 50,000+ applications — under 1%. That’s the same acceptance league as Goldman Sachs (0.7%) and JPMorgan (~0.8%). This is not “be impressive”; it’s lottery-adjacent math even for excellent candidates. The work itself — feature engineering on market data, backtesting, ML infrastructure that has to be fast and correct — screens for competition-math-grade probability and statistics plus strong coding, and the interviews test exactly that. Apply if you have that profile (the expected value is absurd), apply in August the day it opens because the best offices fill first, and do not build your summer plan on it.

Ranges compiled from platform listings, job postings, and worker reports · last verified July 2026.

The actual technical bar (what “requirements” means in practice)

Strip the job postings down and the ML-intern application has three parts.

The coursework floor. Linear algebra, probability and statistics, and at least one real ML course — plus working Python and one framework, PyTorch or TensorFlow, used in anger rather than in a tutorial. That’s the floor, not the differentiator; nearly everyone applying has it.

One shipped project beats a Kaggle-only profile. This is the single most common gap. A profile that’s entirely Kaggle notebooks says “I can optimize a leaderboard on clean data someone else prepared.” A project you scoped, built, and shipped — a fine-tuned model behind a working demo, a merged PR to a known ML repo, an evaluation harness with a write-up of what you found — says “I can do the job.” NVIDIA says out loud what every screener applies quietly: projects and practical skill get weighed alongside GPA, not below it. A strong Kaggle placement is a genuine plus on top of a real project; it just can’t be the whole story. The step-by-step method for manufacturing this kind of proof from zero is in AI jobs with no experience — it’s the same proof-of-work playbook, pointed at ML.

The interview shape. Applied roles run coding rounds plus ML fundamentals — expect to explain bias–variance, regularization, and how you’d debug a model that trains but doesn’t generalize, in plain language. Research roles add a research conversation about work you’ve actually touched. One genuinely new thing for the 2027 cycle: Google and Meta now permit AI assistants during coding interviews, which means the screen is shifting from “can you produce syntax under pressure” toward “can you decompose a problem and check the machine’s work” — practice accordingly rather than grinding memorized solutions.

The Anthropic/OpenAI correction, briefly

Students burn entire application cycles hunting “Anthropic ML internship” or “OpenAI internship.” As of July 2026, those don’t exist as undergrad summer internships. Those labs route early-career people through fellowships and residencies — OpenAI’s Residency requires that you not be enrolled in school, and Anthropic’s programs are full-time cohorts, not summers. The full breakdown of what the labs run instead, with dates, is on the AI internships hub. One pointer worth keeping if you’re a strong ML student who cares about the field’s failure modes: the paid fellowship route into safety-focused research is real money and real research experience — see AI safety jobs for how that path works.

Timing, in one paragraph

It’s July 2026, which means the summer-2027 cycle is about to open, not months away: Amazon opens July–August, Microsoft around mid-August, the quant firms in August (rolling — apply the week they open), NVIDIA August–October, Apple September–November, Google’s STEP window around mid-October for 2–4 weeks, and NSF REU applications November through March. Rolling admission is the default everywhere, and applying within about two weeks of a posting opening materially improves your odds. The full dated, verified calendar — every employer, every window, maintained as dates firm up — lives on the AI internships page; set your alerts from there.

No offer? The ML-specific fallback ladder

Most qualified applicants get zero offers — that’s what sub-1% acceptance and even the ~15–20% conversion estimate at brand-name firms for qualified candidates arithmetically guarantee. The students who come out of a rejected summer ahead are the ones who treat it as a fork, not a verdict. In order of leverage:

1. Research assistant with a professor. Email labs whose papers you’ve actually read, with one specific sentence about their work. Often unpaid or work-study — but it’s the single most reliable route to a recommendation letter and a research line, and it converts directly into REU and Student Researcher applications next cycle.
2. Open-source ML contributions. A merged PR to a known ML repo is verifiable, public, and screams “can work in a real codebase” louder than any certificate. Start with documentation and test fixes to learn the repo, then take a real issue.
3. Paid AI-training work on technical queues. The coding and STEM evaluation queues on legitimate AI-training platforms pay a premium over generalist work precisely because they need people with your coursework — and “evaluated model outputs on coding tasks” is honest, relevant, paid ML-adjacent experience, not resume filler. The vetted platforms and how the work actually functions are in AI training jobs and data annotation jobs.

Stack any one of these with a shipped project and you enter the August 2027 cycle as a different applicant. And if what you actually need this year is income plus a first AI line on the resume, the broader map of roles that hire without internships is in entry-level AI jobs.

Tools that get the interview

The bar above is skill and proof — no tool substitutes for it. But ML intern season is a volume game across a dozen rolling portals, and a few tools save real hours. Our current picks — with the honest caveats and what each actually costs — live on one page: the tools we actually recommend.

FAQ

Do you need a master’s or PhD for a machine learning internship? Depends on the track. Applied ML at big tech and quant internships take undergrads. Research internships are mixed: NSF REU is undergrad-only, DeepMind’s Student Researcher program takes BS through PhD, and Meta FAIR is PhD-only. Check the eligibility line before spending an application — it’s the fastest filter you have.

How hard is it to get a machine learning internship? The honest answer: the famous ones are lottery-adjacent. Jane Street admits under 1% of applicants, and even qualified candidates at brand-name tech firms convert at an estimated 15–20%. The rational play is applying early (within two weeks of opening) and broadly across tiers — including Ai2, MLH, and REU sites — not just FAANG and quant.

Do machine learning interns get paid? Yes, essentially always in the US at legitimate programs. Big-tech ML interns earn roughly $45–62/hour (estimated), NSF REU pays about $6,300 for the summer plus housing, and quant ML interns earn $4,300–6,000/week at the top firms. Anyone charging you a fee to intern is running a scam, full stop.

Can a freshman or sophomore get an ML internship? Rarely a true ML internship — but there are on-ramps built for exactly you: Google STEP and Microsoft Explore target first- and second-years, and the MLH Fellowship is beginner-friendly. Use those years to bank coursework, one shipped project, and a professor relationship; that’s what makes the junior-year ML application land.

Is Kaggle enough to get a machine learning internship? On its own, usually not. Kaggle proves optimization skill on prepared data; screeners also want evidence you can scope and ship — a working demo, a merged PR, an evaluation write-up. A strong Kaggle placement plus one real project is a genuinely strong profile; Kaggle alone is half a profile.

Related guides

• AI internships — the full dated summer-2027 application calendar across every employer, plus what the big AI labs run instead of internships.
• AI jobs with no experience — the proof-of-work method for building the project that gets ML applications past the screen.
• AI training jobs — the paid technical-queue work that doubles as ML experience while you wait out the next cycle.