Rémi Delacourt

Timeline

1. 2024

   Master's in CMU's Catalyst group

   He completes an M.S. in 2024 with the Catalyst group at Carnegie Mellon University. Catalyst is an interdisciplinary machine learning and systems research group that explores problems to automate learning systems.

2. January 2025

   AdaServe speculative decoding system

   In January 2025, AdaServe appears on arXiv with him among its authors. The LLM serving system tailors speculative decoding to each request's latency target, and it is on the program of EuroSys 2026 in Edinburgh, where his affiliation is given as EPFL.

3. June 2025

   The Magistral reasoning model

   In June 2025, he is among the authors of Magistral, Mistral's first reasoning model. The paper describes a scalable reinforcement learning pipeline built on Mistral's own models and infrastructure, with Magistral Medium trained by RL alone on top of Mistral Medium 3 and Magistral Small released as open source under Apache 2.0.

4. July 2025

   Voxtral speech models

   He is among the authors of Voxtral in July 2025, which presents Voxtral Mini and Voxtral Small, two multimodal audio chat models that comprehend spoken audio and text. A 32K context window handles audio files up to 40 minutes, and both models are released under the Apache 2.0 license.

5. 2026

   Co-serving with FlexLLM

   In 2026, FlexLLM appears at USENIX NSDI '26 with him among its authors, listed at Mistral AI. The system co-serves LLM inference and PEFT-based finetuning on shared GPUs by fusing computation at the token level, with end-to-end GPU memory savings of up to 80%.

6. July 2026

   Robostral Navigate for robots

   He is among the authors of Robostral Navigate in July 2026, an 8B vision-language model that navigates robots from a stream of monocular RGB images. It reaches a 77.4% success rate on the R2R-CE benchmark with a single RGB camera.

7. July 2026

   The Shieldstral safety classifier

   Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier, appears on arXiv in July 2026 with him among its authors. The model frames content moderation as a binary question-answering task and matches or outperforms models nearly seven times its size on text safety benchmarks.

8. As of October 2026

   Research engineer at Mistral AI

   He is a Research Engineer at Mistral AI in Paris, with a focus on large-scale machine learning systems. His work covers designing models under hardware- and system-level constraints and then maximizing their speed and efficiency.