From Scratch to SOTA: Training a 3B State-Space Vision Model — Krishna Prasad Srinivasan, Sarvam

Why watch

This video details Sarvam's 3B parameter vision language model trained for multilingual Indian document OCR. It covers the four-stage training curriculum (13T text tokens, 300M image-text pairs, 100M OCR samples, RL), the architectural choice of SSM over transformers for handling 10K+ visual tokens, and the block-level OCR approach that became industry standard. Ideal for ML engineers interested in multilingual VLMs, SSM architecture, and scalable RL with machine-checkable rewards.