Sumukh Aithal

Timeline

1. 2020

   NeurIPS runner-up and Kaggle silver

   With Dhruva Kashyap and Natarajan Subramanyam, he builds a generalization metric based on robustness to augmentations, and it finishes as first runner-up in the NeurIPS 2020 competition on predicting generalization in deep learning. In 2020, he also wins a silver medal, ranked 32 of 1095 teams, in Kaggle's ALASKA2 Image Steganalysis competition with an EfficientNet B7 model.

2. 2021–2022

   Vision and AI Lab at IISc

   He is a research intern for two years at the Vision and AI Lab of the Indian Institute of Science under R. Venkatesh Babu, and coauthors S3VAADA on active domain adaptation at ICCV 2021. In 2022, he coauthors an ICML paper that introduces Smooth Domain Adversarial Training and a NeurIPS paper showing that Sharpness-Aware Minimization lifts accuracy on minority classes by 6.2% over Vector Scaling Loss. He completes his undergraduate studies in computer science and engineering at PES University in 2022.

3. 2022

   Depth cues in contrastive learning

   He works as a research intern at Mila with Anirudh Goyal, collaborating closely with Alex Lamb and Michael Mozer. With Goyal, Lamb, Yoshua Bengio and Mozer, he coauthors "Leveraging the Third Dimension in Contrastive Learning" for a NeurIPS 2022 workshop, where a depth channel raises BYOL's ImageNette accuracy from 85.3% to 88.0%.

4. 2024

   Diffusion hallucinations at Carnegie Mellon

   Having worked as an Applied Researcher II at Fujitsu Research India, he pursues a master's in machine learning at Carnegie Mellon University, advised by Zico Kolter. In 2024, he publishes "Understanding Hallucinations in Diffusion Models through Mode Interpolation" at NeurIPS with Pratyush Maini, Zachary Lipton and Kolter; a variance metric removes over 95% of hallucinations in synthetic datasets while keeping 96% of in-support samples.

5. June 2025

   Magistral, Mistral's first reasoning model

   He coauthors Magistral, Mistral AI's first reasoning model, released in June 2025 together with its own scalable reinforcement learning pipeline. Magistral Medium is trained for reasoning on top of Mistral Medium 3 with reinforcement learning alone, and Magistral Small is open-sourced under Apache 2.0.

6. July 2025

   Voxtral audio chat models

   He coauthors Voxtral, which presents Voxtral Mini and Voxtral Small, two multimodal audio chat models released under the Apache 2.0 license in July 2025. A 32K context window lets them handle audio files up to 40 minutes long.

7. January 2026

   The Ministral 3 model family

   He coauthors Ministral 3, a family of dense language models in 3B, 8B and 14B sizes published in January 2026. Each size comes as a base, an instruction-tuned and a reasoning model, derived through Cascade Distillation and released under Apache 2.0.

8. As of October 2026

   AI Scientist at Mistral AI

   He is an AI Scientist at Mistral AI, where he works on pretraining. His research interest is the science of pretraining and making it more efficient.