Pierre Stock

Pierre Stock is an AI researcher and the VP Science at Mistral AI, where he drives model releases, roadmap strategy and execution. After a master's degree at Paris-Saclay, he began his career at Facebook AI Research in 2017 and finished a PhD there with ENS de Lyon in 2021, with a thesis on efficiency and redundancy in deep learning models. His early papers covered model compression with Quant-Noise in 2020 and the LeViT vision transformer in 2021. He joined Mistral AI as its first employee and co-authored Mistral 7B in 2023, Mixtral 8x7B and Pixtral 12B in 2024, and Magistral, the company's first reasoning model, in 2025. In October 2026, he told TechCrunch that Mistral Large 4 was trained on 4,000 Nvidia GPUs.

Now at: Mistral AI
Works: 37
Badges: Bridge

Publication numbers from OpenAlex, as of 22 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

2017

Joining Facebook AI Research

After a master's degree in Mathematics, Vision, Learning from Paris-Saclay, he begins his career at Facebook AI Research in 2017.
2018–2021

PhD on efficiency and redundancy

He completes a PhD at Facebook AI Research in collaboration with ENS de Lyon, finishing in 2021. Among Rémi Gribonval's doctoral students for 2018–2021, he writes the thesis Efficiency and Redundancy in Deep Learning Models: Theoretical Considerations and Practical Applications.
2020

Quantization noise for compression

In April 2020, he co-authors Training with Quantization Noise for Extreme Model Compression with Angela Fan, Benjamin Graham, Edouard Grave, Rémi Gribonval, Hervé Jégou and Armand Joulin. The method quantizes a different random subset of weights during each forward pass, so that unbiased gradients flow through the other weights.
2021

The LeViT vision transformer

In April 2021, he co-authors LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference with Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Armand Joulin, Hervé Jégou and Matthijs Douze. LeViT is a hybrid neural network for fast inference image classification that introduces an attention bias to integrate positional information.
2023

Mistral 7B at Mistral AI

In October 2023, he is among the authors of Mistral 7B, a 7-billion-parameter language model from Mistral AI, the company he joins as its first employee. The model outperforms Llama 2 13B across all evaluated benchmarks.
2024

Mixtral 8x7B and Pixtral 12B

In January 2024, he is among the authors of Mixtral 8x7B, a sparse mixture-of-experts model in which a router network selects two of eight experts for each token at each layer. In October 2024, he is an author of Pixtral 12B, a 12-billion-parameter multimodal model trained to understand both natural images and documents.
June 2025

Mistral's first reasoning model

In June 2025, he is an author of Magistral, Mistral's first reasoning model, built with the company's own scalable reinforcement learning pipeline. Magistral Medium is trained on top of Mistral Medium 3 with reinforcement learning alone, and Magistral Small is open-sourced under Apache 2.0.
As of October 2026

VP Science at Mistral

He is VP Science at Mistral AI, where he drives model releases, roadmap strategy and execution. In October 2026, Mistral releases Mistral Large 4, a multimodal model with 1 trillion parameters, and he tells TechCrunch it was trained entirely on Mistral's own compute with 4,000 Nvidia GPUs.