Sarvam

Srinivasan frames document intelligence as coherent knowledge extraction: recognizing text is insufficient if its relationships disappear. Complex script encodings and scarce labeled data compound the problem. His account of Sarvam’s three-billion-parameter model separates block-level OCR from layout and reading-order harnesses. The state-space backbone updates a carried state token by token, avoiding the all-pairs attention costs he describes for transformers. With pages producing 5,000–10,000 visual tokens, he argues that linear sequence-processing compute and constant carried-state memory justify some lossy recall. This is an application-specific tradeoff, not a claim that every part of the system uses constant memory.

The reported curriculum starts with 13 trillion text tokens across English, Indian languages, math, and code, followed by 300 million image-text pairs, supervised fine-tuning on 100 million OCR samples, and reinforcement learning. OCR examples include tables, equations, and handwriting. A language prior can help resolve a half-blurred word, but that example explains the training rationale rather than demonstrating a measured recovery result. A later answer supplies the essential qualification: extraction should preserve even mistakes in the original document, rather than rewrite them.

Srinivasan identifies data generation and useful evaluations as central to training languages with little readily available labeled material. He describes synthetic documents, material derived from real documents, and evaluation feedback that improves training inputs. The reinforcement-learning loop samples groups of outputs, scores them with machine-verifiable tests, and reinforces performance relative to the group average. Reward examples include character error rate, table structure, equations, and grammar. He recommends complex real-world documents for useful RL tests; grammar rewards must also respect the stated requirement to preserve source errors. He leaves the quality and usefulness of transliteration-derived training data unresolved.

The speaker reports leading English benchmark results at launch and a continuing lead across 22 Indian languages, while acknowledging that later models have advanced English results. These comparisons are not independently verified by the supplied sources. He describes a human-in-the-loop document workbench with confidence scores, block-level grounding, and agentic proofreading, plus API and on-premises deployment options. His reported digitization workload exceeds 35 million pages across insurance, banking, government, and historical preservation; it is not a stated count of completed pages. A public 22-language benchmark spanning documents from the 1800s onward remains a promised release in the recording, as do broader general-purpose vision models.

Srinivasan makes sovereignty concrete through control over where document data is sent and when and how it is used. The final answers describe a training mix of 40% Indic material, with the remainder English, math, code, and other material. Deployments across different states process regional-language, English, and mixed-language documents. He ends by saying a deployment-feedback pipeline has started, with training on that feedback still upcoming. The ending connects immediate digitization work to future training material without establishing that the feedback has already improved the model.