Research Engineer, Latency & Throughput

About Mistral Mistral provides full-stack AI solutions: from frontier models to developer tools, applications, and compute. We partner with enterprises tackling the hardest problems across high-stakes industries like finance, manufacturing, defense, healthcare, and the public sector, co-creating customized AI systems that they can run on their terms. We are a dynamic, collaborative team passionate about AI and its potential to transform society. Our diverse workforce thrives in competitive environments and is committed to driving innovation. Our teams are distributed between Europe, North America, Asia and the Middle East. We are creative, low-ego and team-spirited. The Role As a Research Engineer on the Latency & Throughput team, you will push the frontier of LLM inference by researching and prototyping techniques that reduce latency and increase throughput — from speculative decoding and quantization to KV compression and new attention variants. The team sits at the intersection of research and production: you will take ideas from recent literature, form hypotheses, run experiments, and ship the techniques that prove out into our serving stack. This role sits between our inference engineers, who own the production serving stack, and our research scientists, who focus on paper-level contributions. You will be more science-focused than a pure engineer: you are expected to generate new ideas, keep up with the state of the art, and implement proof-of-concepts that bridge the gap between academic results and production impact. Some techniques you explore will be lossy or require integration into training — retraining models, tuning architectures, or training auxiliary models to improve inference. You will work closely with the Inference Foundation and Kernels teams, who turn your validated techniques into production recipes, and with pretraining and RL teams when your work touches model training. Success is measured by demonstrated latency or throughput gains, first at small scale, then adopted in production. What You Will Do • Research and prototype inference techniques — speculative decoding, quantization, KV compression, watermarking, and new model architectures — by reading recent literature and forming testable hypotheses. • Design and run experiments to validate or refute ideas, iterating quickly on small-scale setups before proposing production integration. Demonstrate clear latency or throughput gains without unacceptable quality loss. • Collaborate with the Inference Foundation team to integrate validated techniques into the vLLM-based serving stack. • Partner with pretraining and RL teams when techniques require co-training, retraining, or training auxiliary models to improve inference. Build the right tooling and validation platforms to facilitate AI-driven inference research. What We're Looking For • A strong ML background with solid knowledge of LLM training techniques — architecture, tokenization, fine-tuning, and evaluation. • Hands-on experience with Python and PyTorch; you write clean, experimental code and iterate fast. JAX is acceptable with strong Python proficiency. C++ and Rust are not required, CUDA and GPU knowledge is expected but hands-on experience is not required. • Research experience of some form: a thesis, publications, a dedicated research project, or efficiency-oriented work in a lab or industry research setting. A PhD is not required, but we are looking for genuine research depth. • A track record of conducting experiments: forming hypotheses, designing controlled comparisons, and drawing clear conclusions from results. • Familiarity with the state of the art in LLM inference — you know what has been tried, what works, and where the open problems are. • Self-starter, low-ego, and collab