How to become a machine learning engineer in 2026: a roadmap built from real job ads

To become a machine learning engineer, learn Python, core machine learning and deep learning with PyTorch, then prove you can put a model into production and keep it working. Employers care more about that proof than about a specific diploma. For most people, the path runs through a software or data job first, or through an internship.

We read the 49 machine learning engineer roles live on Tokn as of 12 October 2026, at 21 European companies. Only 7 of the 34 ads with a requirements list mention a degree; more than half ask for ML shipped to production. For the wider picture, see the eight paths into an AI career.

In short

• ML engineers build, deploy and monitor models inside products, often including LLMs now.
• Degrees matter less than you think: 5 of the 7 ads that mention one accept experience instead, and none requires a PhD outright.
• The core stack is Python (33 of 49 ads) and PyTorch (31), plus production experience.
• We estimate 8 to 14 months of part-time study, but most ads also want 2 to 5+ years of experience: plan a first step through software, data or an internship.
• Paris and London lead, then fully remote roles. More than a third of the roles are at product companies, not AI labs.

What does a machine learning engineer do?

A machine learning engineer turns models into features that real users rely on. Pennylane summed it up in an ad: "own the full lifecycle of your solutions: problem framing, data and labeling strategy, baselines, training, evaluation, deployment, experimentation, monitoring".

A typical week can include:

• Deciding with product managers whether to build a model at all. Photoroom asked candidates to "determine when to fine-tune or build a model yourself versus integrating an external model, API or existing library".
• Training and debugging models, which Wayve described as "deep work / coding".
• Running experiments. Spotify wanted its messaging ML engineer to "plan and run A/B experiments in a multi-objective environment".
• Meeting production constraints. A Doctolib retrieval role meant systems "handling 100M+ documents, 50+ req/s, and sub-300ms latency requirements".

The job now overlaps heavily with LLM work: at least 19 of the 34 ads with requirements mention LLMs, generative AI, RAG, retrieval or agents. LLM engineering deserves its own guide, so we leave it aside.

ML engineer vs data scientist vs AI engineer

ML engineers put models into products; data scientists produce insights; research engineers develop new methods; AI engineers mostly build on existing models. As Chip Huyen writes in her book on machine learning interviews, "data scientists tend to be better statisticians, and ML engineers tend to be better engineers."

The titles are blurring. On Tokn, "AI engineer" appears in 59 active job titles against 34 for "machine learning engineer". LinkedIn's Jobs on the Rise 2026 ranked AI engineer, not ML engineer, as the fastest-growing job in France, Italy, Spain and the UK, according to Euronews. Search for both titles.

What employers actually ask for in 2026

Employers ask for Python, PyTorch and, above all, proof that you have shipped ML to production. Most also want several years of experience.

Our method: every active Tokn job titled machine learning engineer, ML engineer, applied scientist or ML research engineer, plus ML internships, excluding managers and ML infrastructure. That gives 49 roles, 34 with a requirements list. Tokn stores shortened requirements, so theme counts are minimums.

Experience: mostly a second job

About half the roles are senior or above: 14 senior and 12 lead, staff or principal, against 3 mid-level, 8 junior and 12 with no level stated. Only 12 ads state a minimum number of years, a small sample, and their median is 5. Of the 8 junior-tagged roles, 6 are internships and the other 2 ask for 2 years: no permanent role in the set asks for less.

Skills: Python and PyTorch first

PyTorch beats TensorFlow almost 3 to 1, so learn it first. For the board-wide view, see the tools AI job postings ask for.

The themes that keep coming back

Six themes stand out in the 34 requirement lists:

• LLMs, RAG, retrieval or agents (at least 19): Doctolib, Doctrine, Ubisoft, Owkin, Mistral AI, Perplexity.
• Production (at least 18): "You've shipped ML into production" (Photoroom).
• A named deep learning framework (at least 15): PyTorch, TensorFlow or JAX.
• Scale and GPUs (at least 12): distributed training, large-scale systems, CUDA.
• Search, ranking and recommendation (at least 11): Perplexity, Spotify, Doctrine, Doctolib.
• Evaluation and experiments (at least 11): Datadog wanted someone who "can define what 'good' means before you measure it".

Classical ML has not gone away. Qonto wanted engineers who know when GenAI fits and when proven ML techniques do better, and Voodoo listed "scikit-learn, LightGBM, PyTorch".

Do you need a degree or a PhD to be an ML engineer?

You don't need a PhD, and often not a specific degree. Only 7 of the 34 ads with requirements mention a degree, and 5 of those 7 accept experience instead, like Datadog's "BS/MS/PhD … or equivalent experience". No ad requires a PhD outright.

The two exceptions: Owkin required a "Master's or Ph.D. in Computer Science, Data Science, or a related quantitative field", and Mistral AI's forward-deployed ML internship wanted students "from a tier 1 engineering school or university".

A PhD helps in research-flavoured roles. Perplexity's ML research engineer ad wanted a "strong publication record" at conferences such as NeurIPS or ICML, and publications appear in 4 of the 34 ads. If that is your goal, browse research engineer and scientist roles.

One caveat, in our view: a strong CS, maths or engineering degree still opens most internship doors. Without one, the usual route is a software job first.

How to become a machine learning engineer: a realistic roadmap

Plan on about 8 to 14 months at 10 to 15 hours a week, roughly half that full-time. These are our estimates, not promises; popular guides claim 6 to 9 months (DataCamp) or 6 to 12 months (Zero To Mastery). Each phase answers something the ads ask for.

Resources worth your time (as of October 2026)

Certifications are optional. The Google Cloud Professional Machine Learning Engineer exam costs $200 and recommends 3+ years of industry experience. The AWS Machine Learning – Specialty is retired: its last exam date was 31 March 2026. We think a shipped project is worth more than either.

Projects that prove you can do the job

Good projects mirror what the ads ask for and show judgement, not a finished tutorial:

1. Ship one model end to end: data, training, evaluation, API, monitoring, public demo. It answers Photoroom's "You've shipped ML into production."
2. Build a search or recommendation system with offline metrics. Perplexity's search ML internship asked for "projects related to information retrieval, ranking, or recommender systems."
3. Fine-tune and evaluate an open model on an eval set you wrote. Figma asked for "experience curating and developing eval sets and LLMs as judges".
4. Make training faster or cheaper with profiling, mixed precision or several GPUs. Datadog wanted people who can "make models run efficiently within real hardware constraints".
5. Contribute to an open-source ML library. Hugging Face asked for "merged pull requests to ML or data libraries, that we can review on GitHub".

Write each up with numbers: baseline, result, latency or cost, next steps.

How people get into ML engineering: four routes

Most ML engineers come from software engineering, data science or an internship. Direct junior hires are rare.

From software engineering

The most direct route: you already have production habits. Mistral AI's forward-deployed ML engineer roles asked for "2+ years as a technical individual contributor (data scientist or software engineer) on AI-based products". Product companies such as Doctolib, Qonto and Pennylane hire ML engineers into product teams, so internal moves are realistic.

From data science

Add software engineering and deployment. Qonto wanted someone who had "developed and deployed client-facing ML products end-to-end — not internal tools or dashboards".

Through an internship

In October, ML internships were advertised by Cohere (fully remote), Perplexity (London and Berlin), Mistral AI (two in Paris) and Ekimetrics (Paris). Doctolib, Shift Technology, Artefact and Together AI advertised adjacent internships, and H Company and Helsing PhD research internships. Tokn listed no apprenticeships.

Straight into a junior role

Rare. Only 2 permanent roles in the set were tagged junior, at Photoroom and Synthesia, and both asked for 2 years of experience.

For CVs, applications and interviews, see our step-by-step guide to landing an AI job.

Where are the ML engineer jobs in Europe?

Paris and London lead by a wide margin, followed by fully remote roles and Berlin. Six of the 49 were fully remote, at Hugging Face, Pennylane, Qonto, Synthesia and Cohere (an internship); see more remote AI jobs.

ML engineering is not only a lab job: 18 of the 49 roles (37%) were at companies that are not AI-native, such as Ubisoft, Qonto and Figma. For the bigger list, see the AI companies hiring the most in Europe.

None of the 49 ads showed a salary we could read. On Levels.fyi, median self-reported total pay for ML engineers was about £112k in London and €82k in Paris in October 2026; see what ML engineers earn across Europe.

Find machine learning engineer jobs on Tokn

Browse all machine learning jobs; the category also includes applied AI and solutions roles, so read titles carefully. Then set up a job alert for ML engineer roles in your city or for remote work.

FAQ

Can you become a machine learning engineer without a degree?

Yes, but rarely as a first job. On Tokn, only 7 of 34 ML engineer ads with requirements mention a degree, and 5 of those accept equivalent experience. Employers want evidence of shipped ML, so the usual route is software engineering first.

Do you need a PhD to be an ML engineer?

No. No ML engineer ad on Tokn required one outright as of 12 October 2026. A PhD or publications matter mainly for research-flavoured roles at labs such as Perplexity and Mistral AI.

How long does it take to become an ML engineer?

We estimate 8 to 14 months of part-time study, about half that full-time. Most ads also want 2 to 5+ years of experience, so the full path usually takes 1 to 3 years, often via a software or data job.

What maths do you need for machine learning engineering?

Linear algebra, probability and statistics, and multivariable calculus, the level Stanford's CS229 lists as prerequisites. Only 5 of 34 ads name maths, but interviews and the work assume it.

Can a software engineer become an ML engineer?

Yes, and it is the most direct route: add core ML, PyTorch and evaluation to the production skills you have. Mistral AI's forward-deployed ML roles accepted 2+ years as a software engineer or data scientist.