← All speakers

Devendra Singh Chaplot

Conference affiliation: Mistral AI · 2024

Devendra Singh Chaplot is an AI researcher and adviser to Sarvam who helped develop Mistral AI’s early open-weight language models, led its multimodal research, and worked on model-training infrastructure at Thinking Machines Lab. His work spans robots finding objects in unfamiliar homes, models processing images at their native resolution, and tools that let developers customize models without operating distributed training infrastructure.

On this page

Mapping unfamiliar environments

Devendra graduated from IIT Bombay in 2014 after studying computer science and engineering, then completed a master’s degree in language technologies and a doctorate in machine learning at Carnegie Mellon. His doctoral research examined autonomous navigation: how a robot can combine perception, spatial memory, exploration, and planning to move through an unfamiliar environment.

His collaborative Active Neural SLAM work separated learned mapping and exploration from conventional planning. A robot builds a spatial representation, chooses where to explore, and uses a planner to reach that destination. The subsequent object-goal navigation system incorporated semantic information so the robot could search for a particular kind of object, such as a chair. The team’s system won the CVPR 2020 Habitat ObjectNav Challenge.

After completing his doctorate in 2021, Devendra worked as a research scientist at Facebook AI Research. His collaborative research in real homes investigated how navigation systems transfer from simulation to unfamiliar physical environments. Explicit maps and modular planning helped connect learned perception to movement beyond the conditions encountered during training.

Building Mistral’s language and vision models

Devendra joined Mistral AI’s founding team in 2023. He coauthored the Mistral 7B and Mixtral papers, helped train Mistral Large, and established the company’s Palo Alto office. He subsequently led the multimodal research team behind Pixtral 12B and Pixtral Large. Pixtral’s architecture processes images at their native resolution and aspect ratio while retaining text capabilities, accommodating visual inputs of different shapes rather than requiring a single fixed image format.

In his 2024 talk on Mistral’s models, Devendra explained several decisions behind the company’s approach:

  • Separate model capacity from active computation. Mixtral 8x7B contains approximately 47 billion parameters but uses roughly 13 billion per token, selecting two of eight experts at each layer. This lets the model draw on a larger parameter pool while limiting the computation required for each prediction. Devendra emphasized performance relative to inference cost rather than total parameter count alone.
  • Let open releases support a commercial business. Public weights give developers room to customize models and deploy them privately. Community experiments, including running models on laptops and phones, also teach the company about deployment possibilities it has not built itself. Devendra described open releases as a source of product awareness and customer acquisition: developers can start with open models and later pay for hosted services or stronger commercial offerings.
  • Prototype broadly, then specialize around a known task. Devendra recommended starting with a powerful general-purpose model while establishing what an application needs. Once the workload and volume are clear, task-specific fine-tuning of an open model can improve the performance-to-cost ratio. Private deployment and control over data provide additional reasons to choose accessible weights.

He also cautioned against treating more training data as an automatic improvement. Noisy data can hurt performance; cleaning, deduplication, and dataset composition require substantial work. The cost of large training runs makes those choices consequential, while settings that work for a smaller model may fail to transfer to a larger one.

Training tools and subsequent roles

In 2025, Devendra joined Thinking Machines Lab’s founding team and worked on Tinker. The API exposes core model-training operations while managing the distributed infrastructure needed to execute them, allowing researchers to customize models without running that infrastructure themselves.

In March 2026, he announced that he was joining SpaceX and xAI. Later that year, he became a part-time adviser to Sarvam, supporting frontier-model development for India. His advisory position includes advocating domestic model-training expertise, computing infrastructure, and technical talent so systems can address local languages and requirements.