Sandeep Subramanian

Also known as MaximumEntropy

Sandeep Subramanian is a language-model researcher at Mistral AI who co-authored the Mixtral, Pixtral, Magistral and Ministral 3 papers. He earned a master's in Language Technologies at Carnegie Mellon in 2017 and a PhD at the University of Montreal, where Chris Pal and Yoshua Bengio advised him. In 2018, he was first author of a multi-task framework for sentence representations trained on over 100 million sentences. As a research manager at NVIDIA, he co-authored the 2023 post introducing the NeMo Data Curator, a tool for preparing trillion-token pretraining datasets. At Mistral, he was an author of Mixtral of Experts and Pixtral 12B in 2024, of Magistral, the company's first reasoning model, in 2025, and of Ministral 3 in 2026.

Now at: Mistral AI
Badges: Bridge

Timeline

2017

Master's at Carnegie Mellon LTI

A research assistant in Carnegie Mellon's Language Technologies Institute, he works with advisor Madhavi Ganapathiraju and helps build Swych, a language-learning app based on code-switching. He completes a Master of Language Technologies there in 2017.
2017–2018

MILABOT and sentence encoders

He becomes a PhD student at Mila (Université de Montréal), advised by Chris Pal and Yoshua Bengio. In 2017, he is among the authors of MILABOT, a deep reinforcement learning chatbot that the Montreal Institute for Learning Algorithms builds for the Amazon Alexa Prize. In 2018, he is first author, with Adam Trischler, Yoshua Bengio and Christopher Pal, of a multi-task framework that trains one recurrent sentence encoder on over 100 million sentences.
August 2023

NeMo Data Curator at NVIDIA

A research manager at NVIDIA with a Ph.D. in deep learning from the University of Montreal, he co-authors the August 2023 technical post that introduces the NeMo Data Curator. The tool curates trillion-token multilingual datasets for LLM pretraining, and his own research there centers on large language model training and alignment.
January 2024

Mixtral of Experts at Mistral

He moves from NVIDIA to Mistral AI and is among the authors of Mixtral of Experts, posted in January 2024. The paper describes Mixtral 8x7B, a sparse mixture-of-experts model whose router selects two of eight experts for every token, released under the Apache 2.0 license.
October 2024

Pixtral 12B multimodal model

In October 2024, he is among the authors of Pixtral 12B, a 12-billion-parameter multimodal model that understands natural images and documents. It uses a vision encoder trained from scratch, handles a 128K-token context window, and is released under Apache 2.0.
June 2025

Magistral reasoning model

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

The Ministral 3 family

The Ministral 3 report, posted in January 2026, lists him among its authors. It presents dense language models in 3B, 8B and 14B sizes for compute and memory constrained applications, each released as a base, an instruction-tuned and a reasoning variant under Apache 2.0.
As of October 2026

Researcher at Mistral AI

He is a researcher at Mistral AI, which he joined after working at NVIDIA.