Sarvam Launches Vision 2.1 With Improved Document Intelligence Across 22 Indian Languages

AI startup Sarvam AI has launched Sarvam Vision 2.1, a document intelligence model with new capabilities for structured extraction, complex table parsing, and Indic handwritten text recognition. The company said the model achieves state-of-the-art results on its Indic OCR benchmark and the olmOCR-Bench, which evaluates how vision models convert complex PDF files into Markdown format.

Sarvam Vision 2.1 is available through Sarvam’s document intelligence APIs, including Digitize API for converting multi-page documents, tables, and handwritten content into structured text, and Extract API for extracting key-value pairs, tables, and form fields.

The company said the new model addresses gaps identified in the earlier version of Sarvam Vision launched in February this year, particularly in multi-page table extraction, key-value extraction from forms, and Indic handwriting recognition. It also includes changes to reduce hallucinations and inconsistencies in the model.

“Sarvam Vision 2.1 is a step change in terms of its new and improved capabilities,” the company said in a statement. “It achieves frontier performance on benchmarks, and is built for production workflows.”

On the official olmOCR-Bench, Sarvam Vision 2.1 scored 87.30, ahead of Infinity-Parser2 Pro at 86.10 and Opus 5 at 85.10. Sarvam Vision 1.0 scored 84.10 on the same benchmark.

On OmniDocBench v1.6, which evaluates text, table structure, and formula recognition, Sarvam Vision 2.1 scored 94.97. PaddleOCR-VL 1.6 led the benchmark with 96.01, followed by Sarvam Vision 2.1 and GLM-OCR at 94.71.

Sarvam also introduced an Indic OCR benchmark covering 22 Indian languages. It evaluates character and word accuracy using material from newspapers, brochures, textbooks, and historical writings dating from 1800 to the present.

Sarvam Vision 2.1 recorded 87.39% overall word accuracy on the benchmark, compared with 84.94% for Bodhan Indic-OCR and 79.35% for Gemini 3.6 Flash. Google Cloud Vision scored 71.76%, while GPT 6 Astra recorded 63.69%.

The company trained the model on both synthetic and real-world data for key-value extraction and Indic handwriting. Sarvam also carried out supervised fine-tuning followed by reinforcement learning with verifiable rewards, or RLVR.

Sarvam said it has also optimised the inference stack to reduce the cost of serving the model. The API has been optimised for production workloads following feedback from users of the first version.