Senior AI Researcher, Pre-training
Hybrid · Heidelberg (Germany) · Senior · Full-time
Want offers like this one?
Create your free PyHires profile: you will see how well you match each offer and get a weekly email with the ones that fit you.
Overview
- Work mode
- Hybrid
- Location
- Heidelberg (Germany)
- Engagement
- Full-time
- Seniority
- Senior
- Languages
- English
- Open to candidates from
- EU countries only
- Company
- Alephalpha
- Posted
- Nov 19, 2026
About the role
Aleph Alpha pre-trains its own foundation models, such as the recently released Kolibri-1, for customers in finance, manufacturing and public administration who need models that understand German and meet European regulation. As a Senior AI Researcher in the Pre-training team you will own core elements of the training recipe and shape the architecture of the next generation of models, working on site in Heidelberg regularly.
What you'll do
- Design and run ablations on the training recipe (optimizers, schedules, initialization) and implement PyTorch-based architectural changes, measuring their effect on convergence, stability and efficiency
- Help monitor large-scale training runs and investigate convergence issues such as loss spikes and divergence
- Track down distributed training failures with the team
- Work with the Compute Performance, Data, Evaluation and Post-Training teams on how hardware and the model lifecycle shape pre-training decisions
- Pair with experienced researchers on live problems from your first weeks
What you bring
- Proficiency in Python and deep familiarity with PyTorch-based training workflows
- The right to work in Germany and regular on-site collaboration in Heidelberg
- An understanding of how transformers work, having pre-trained LLMs from scratch yourself, even at small scale
- A track record of going deep into unfamiliar technical areas, such as reimplementing papers or debugging far outside your comfort zone
- Clear communication and openness to feedback
Nice to have
- Scaling laws, hyperparameter transfer (such as muP) or recent work on optimizers and training stability
- Publications at NeurIPS, ICML or ICLR, or contributions to open-source ML training libraries
What they offer
- Hands-on time on large-scale clusters with thousands of NVIDIA Blackwell GPUs
- A flat structure with ownership, autonomy and pair or mob programming on hard problems
- Flexible working hours, a hybrid working model and a subsidised Germany-wide transport ticket
Requirements
Based on:Python
Must-have
- LLM
- PyTorchEssential
About Alephalpha
Aleph Alpha is one of the few European companies pre-training its own foundation models, such as Kolibri-1, for customers in finance, manufacturing and public administration who need models that understand German, meet European regulatory requirements and work reliably in high-stakes settings. Its pre-training team is based in Heidelberg.
Stack
- Python
- LLM
- PyTorch
Company
Alephalpha
Germany