Senior Research Scientist | Model Scaling

About the role

DeepL is seeking a Senior Research Scientist to guide the core modeling decisions for our next generation of translation models. This is a hands-on position where you will rapidly prototype, conduct large-scale experiments, and transition modeling choices into production. Your work will directly impact the foundational architecture of our most advanced language AI systems. You will act as a key technical driver in defining how models grow smarter and more efficient. This role requires a strong partnership with engineering to ensure research breakthroughs translate into robust products. You will challenge existing scaling laws and propose novel architectures for multilingual understanding. The position demands intellectual rigor and a focus on measurable improvements in real-world scenarios.

Key facts

What you'll do

• Select and evaluate open foundation models as the basis for future translation systems.
• Make key architectural decisions for scaling models to hundreds of billions of parameters, including sparse and efficient designs like Mixture-of-Experts.
• Develop multi-capability adaptation strategies using methods such as LoRA and PEFT.
• Manage the full modeling lifecycle, from prototyping and experiments to evaluation and production delivery, ensuring rigorous and reproducible results.
• Collaborate with specialists in post-training, reinforcement learning, and instruction-following to integrate alignment and capabilities into base models.
• Monitor advancements in open-model and scaling research, providing informed recommendations to the team.
• Design and execute large-scale experiments to validate hypotheses about model behavior and efficiency.
• Analyze performance metrics across languages and domains to identify systemic gaps.
• Partner with data teams to curate and improve training corpora for scaling initiatives.
• Implement debugging strategies for complex training pipelines to resolve instability or degradation.
• Translate ambiguous product requirements into concrete modeling experiments and success criteria.
• Document methodologies and findings to support knowledge sharing and long-term maintenance.
• Mentor junior researchers on best practices for model evaluation and experimental design.
• Participate in code reviews to ensure high standards of software engineering in research.

Requirements

• Extensive practical experience adapting and scaling large language models (multi-billion parameters) beyond basic usage, through fine-tuning, instruction-tuning, or post-training.
• Strong judgment regarding architectural trade-offs at scale (e.g., dense vs. MoE) and selecting appropriate open-weight foundation models.
• Functional knowledge of parameter-efficient and multi-capability adaptation techniques (LoRA/PEFT and variations).
• A hands-on approach to building, training models, running experiments, and debugging pipelines, capable of bringing research to production with engineering teams.
• Strong coding and experimentation abilities.
• Clear communication skills, ability to collaborate across teams, and align research with product and engineering goals.
• Demonstrated ability to work in a fast-paced environment with evolving priorities.
• Experience working with multilingual text data and understanding linguistic diversity challenges.
• Commitment to writing clean, maintainable code that supports long-term research goals.
• Willingness to engage in deep technical discussions with senior leadership on strategic direction.

Nice to have

• Experience quantifying uncertainty in large models through methods like Bayesian approaches, ensembling, steering, or prompt-based techniques.
• Background in machine translation, multilingual NLP, or document/layout-aware modeling.
• Familiarity with MoE-specific training and adaptation (e.g., expert routing, Mixture-of-LoRA-Experts) and large-scale data-mixture design.

Skills & tools

• Python
• PyTorch
• JAX
• Tensorflow

Practical notes

DeepL offers a hybrid work schedule with two office days per week. Employees receive Virtual Shares, linking their contributions to company growth. Benefits include 30 days of annual leave (plus public holidays) and access to mental health resources. The company hosts monthly full-day "Hack Fridays" for personal projects and cross-team collaboration. This role is based in London and requires compliance with local working regulations. Travel may be required for team meetings or industry events as needed. The position is full-time and expects availability during standard business hours for collaboration. Candidates must be authorized to work in the United Kingdom without sponsorship for this role.