Research Engineer, GPU Performance & Scale

Mistral.ai; Paris; On-site; EUR 110,000 - 160,000; Full time

Benefits offered by this job

Healthcare coverage

Parental leave

Relocation support

Wellness programs

Meal and transportation allowances

Job summary

Mistral AI is seeking a Research Engineer, GPU performance, to work at the core of our frontier AI models. You will write low-level code to extract maximum GPU throughput, collaborating with Research Scientists to push architectures toward efficient inference and scalable training.

You’ll integrate improvements into our MLOps framework, profile on thousands of GPUs, and help shape how hardware resources are measured and served for large language models.

Qualifications

• Experience writing custom low-level kernels for GPUs.
• Familiarity with distributed GPU clusters and high-performance computing.
• Strong knowledge of generative AI models and fine-tuning.
• Software design basics: testing, code review, CI/CD.
• Self-motivated, collaborative attitude.

Responsibilities

• Develop low-level code to maximize GPU capacity, from custom kernels to memory layout and communication optimisations.
• Rethink model architectures for efficient inference with Research Scientists.
• Integrate low-level optimisations into a high-level MLOps framework for broad impact.
• Profile, benchmark, and optimize performance across thousands of GPUs during training and inference.
• Assess hardware usage and build serving performance estimates for LLM execution.

Skills

Low-level GPU kernels

Distributed GPU clusters

Generative AI understanding

CI/CD and testing

Self-starter

Job description

Mistral AI is seeking a Research Engineer, GPU performance, to work at the core of our frontier AI models. You will write low-level code to extract maximum GPU throughput, collaborating with Research Scientists to push architectures toward efficient inference and scalable training.

You’ll integrate improvements into our MLOps framework, profile on thousands of GPUs, and help shape how hardware resources are measured and served for large language models.