Devendra Singh Chaplot is an artificial-intelligence researcher and adviser to Sarvam who helped develop Mistral AI’s early language models, led its multimodal research, and joined the founding team at Thinking Machines Lab. His contributions span autonomous robotics, open-weight language models, multimodal systems, and infrastructure for training customizable AI.
From autonomous navigation to frontier AI
Chaplot studied computer science and engineering at IIT Bombay, graduating in 2014, before completing a master’s degree in language technologies and a doctorate in machine learning at Carnegie Mellon. At Facebook AI Research, he investigated how robots understand unfamiliar environments and navigate toward specific goals. His Object Goal Navigation system combined semantic mapping, exploration, and classical planning, winning the CVPR 2020 Habitat ObjectNav Challenge. Subsequent research in real homes showed how modular navigation systems can transfer more reliably from simulation into physical environments.
He joined Mistral AI’s founding team in 2023, coauthoring the Mistral 7B and Mixtral papers and establishing its Palo Alto office. He subsequently led the multimodal team behind Pixtral 12B and Pixtral Large; Pixtral’s architecture processes images at their natural resolution and aspect ratio while maintaining strong text capabilities.
At Thinking Machines Lab, Chaplot worked on Tinker, a developer-facing model-training API that exposes core training operations while managing infrastructure. In March 2026, he announced that he was joining SpaceX and xAI. He subsequently became a part-time adviser to Sarvam, supporting frontier-model development for India.
- Modular real-world navigation: Explicit semantic maps and conventional planning help robots overcome the simulation-to-reality gap.
- Sparse mixture-of-experts efficiency: Mixtral 8x7B contains approximately 47 billion parameters but activates roughly 13 billion per token, lowering inference costs without requiring every parameter for each prediction.
- Open-weight models as practical infrastructure: Open distribution supports customization, private deployment, community adoption, and commercial upgrades. Chaplot recommends prototyping with powerful general-purpose models, then fine-tuning an open model when production costs, privacy, or control become decisive.
- Accessible model-building capability: Pixtral, Tinker, and his advisory work reflect an emphasis on systems that developers can adapt to particular visual tasks, languages, and deployment requirements.
