Senior LLM Research Engineer - Training, Post-Training & Model Efficiency

We are recruiting for a well-funded deep-tech company developing more efficient and capable AI models.

Joining its Research team, you will work directly on large language and multimodal models, taking experiments from initial ideas through training, evaluation and efficient inference. The work spans distributed training, reinforcement learning, model compression and distillation, with opportunities to contribute to research papers, open-source code and model releases.

This role suits someone who enjoys working inside models: understanding training behaviour, changing implementations, debugging difficult problems and building evidence through rigorous experiments.

The role

Depending on your strengths and research priorities, you will:

- Train and post-train language and multimodal models on multi-node GPU clusters, using supervised fine-tuning, preference optimisation, reinforcement learning and distillation.
- Develop and evaluate compression techniques, including quantisation, pruning and low-rank methods, alongside recovery training to preserve model capabilities.
- Build reward functions, verifiers and training environments for reasoning, coding and agentic tasks.
- Investigate alternative and hybrid architectures, including state-space models, linear attention and mixture-of-experts.
- Develop data pipelines for curation, filtering, deduplication and synthetic data generation.
- Design reproducible experiments and evaluations covering reasoning, coding, long-context performance and instruction following.
- Profile and improve inference performance across GPUs and other accelerators, assessing latency, memory use and throughput.
- Share findings, review code and mentor colleagues.





What you will bring

- At least four years of hands-on experience training or post-training transformer models.
- Strong PyTorch skills, including experience modifying model implementations, custom losses and training configurations.
- Experience running distributed training across multiple nodes using frameworks such as FSDP, DeepSpeed or Megatron-LM.
- Practical understanding of parallelism strategies and cluster environments using Slurm or Kubernetes.
- Hands-on experience with supervised fine-tuning, DPO and at least one reinforcement-learning method, such as PPO or GRPO.
- Strong foundations in linear algebra, optimisation, probability and training dynamics.
- An understanding of GPU memory, mixed precision, profiling and performance bottlenecks.
- Rigorous experimental practice, including ablations, reproducibility, experiment tracking and careful evaluation.
- A Master’s or PhD in a relevant subject, or equivalent industry experience.
- Strong debugging and communication skills, with fluent English.

Additional experience we would welcome You do not need experience in every area. Depth in some of the following would be valuable:

- LLM post-training frameworks such as TRL, verl, OpenRLHF or NVIDIA NeMo.
- Quantisation-aware training, pruning, distillation or other model-compression techniques.
- RL rollout infrastructure, reward modelling, verifiers and agentic environments.
- Multimodal training, world models or vision-language-action models.
- State-space models, linear attention or other emerging architectures.
- Inference optimisation, CUDA or Triton kernels, and deployment on alternative accelerators or edge devices.
- Research publications, substantial open-source contributions or publicly released models.

📌 Senior LLM Reseach Engineer (Barcelona)
🏢 DeepRec.ai
📍 Barcelona