Markdown source

From Scratch to SOTA: Training a 3B State-Space Vision Model for 1.4 Billion People

Conference Context

Session Description

India has 22 official languages. Across those languages live over a billion people whose knowledge is locked inside scanned images in scripts that most frontier models perform poorly. The problem is dire - until now, there wasn't even a comprehensive benchmark to measure Indic OCR performance, let alone training data at scale. When Sarvam AI set out to solve this, we had to build the infrastructure before the model, creating the first ground-truth benchmark for Indic document intelligence. In this talk, Krishna Srinivasan, who led the Vision Models team to build India's first sovereign VLM from scratch, will walk through the end-to-end engineering lifecycle. We will cover: (a) Architecture: Why we chose a 3B-parameter state-space architecture over transformer baselines to handle high-resolution visual inputs with minimal memory overhead and faster inference. (b) Training Pipeline: The exact recipe we used: starting with text-only pre-training, moving to continual pre-training with text and images, followed by SFT. Finally, we'll cover the advances we made in implementing large-scale RL with Verifiable Rewards for visual tasks in just 3 days using deterministic character-level reward signals. (c) Compute Efficiency: How we trained a frontier-competitive multimodal model with extreme capital efficiency, optimizing distributed training and GPU cluster management to punch far above our compute class. (d) Agentic Workflows: How this model powers Sarvam Akshar, a first-of-its-kind agentic document intelligence workbench featuring visual grounding and automated proofreading loops. The results speak for themselves: Sarvam Vision achieves best-in-class global scores (84.3% on olmOCR-Bench, 93.28% on OmniDocBench) and dominates Indic OCR. Attendees will learn the blueprint for compute-efficient multimodal training, and deploying state-space VLMs for population-scale enterprise workloads.

Media Evidence

No related AI Engineer channel video found yet.

Evidence Graph

This evidence graph is generated from currently linked source material: official schedule text, related video pages, cached transcripts, visible slide text, dense/reconstructed slide pages, and AI slide-classification audits.

Media Signals

No linked video, transcript, or slide source has been attached yet.

Agent Reading Notes

Use these signals to refine the synopsis, topic links, people/company context, and method notes. If a source is a related external video rather than an exact official recording, keep it framed as supporting evidence.

Transcript Status

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

People

Notes

Synthesis

Synthesized Breakdown

From Scratch to SOTA: Training a 3B State-Space Vision Model for 1.4 Billion People ## Conference Context - Date/time: 2026-06-29 · 3:20pm-3:40pm - Track/room: Vision & OCR · Track 2 - Speaker(s): Krishna Prasad Srinivasan - Session type/status: sponsor · confirmed - Track: Vision & OCR - Room: Track 2 - Session type: sponsor - Status: confirmed ## Session Description India has 22 official languages. Across those languages live over a billion people whose knowledge is locked inside scanned images in scripts that most frontier models perform poorly. The problem is dire - until now, there wasn't even a comprehensive benchmark to measure Indic OCR performance, let alone training data at scale. When Sarvam AI set out to solve this, we had to build the infrastructure before the model, creating the first ground-truth benchmark for Indic document intelligence.

Speaker And Company Context

Topics Covered

Derived Links And Source Material

Novel Concepts / Clever Methods

Evidence Boundary

This synthesis is based on the official schedule and linked source pages. It should be revisited when exact session recordings or transcript-backed secondary sources are available.