Illustration by tuput
English
On 24 September 2026 Sarvam released Vision 2.1, a document-reading model for English and all 22 scheduled Indian languages that now extracts fields from forms, parses multi-page tables and reads Indic handwriting. Sarvam reports 87.3 on olmOCR-Bench and 87.39 on its own Indic test, and every score so far is Sarvam's.
On 24 September 2026 Sarvam, an AI company in Bengaluru, released Sarvam Vision 2.1, a model that reads scanned and photographed documents in English and all 22 scheduled Indian languages. Three abilities are new: pulling named fields out of forms, parsing tables that run across several pages, and reading handwriting in Indian languages. Sarvam reports 87.3 on olmOCR-Bench, a public English-language test, and 87.39 on a new Indic test it built itself. Every score in this article comes from Sarvam, and no independent re-run had turned up as of 12 November 2026.
The task is optical character recognition, usually shortened to OCR: turning a picture of text into text that a computer can search, copy and check. Sarvam’s February 2026 launch post says much of India’s knowledge sits in physical documents, scanned archives and historical collections.
What Vision 2.1 adds
Sarvam’s post names three new capabilities: key-value extraction from forms (the label and the entry beside it, such as “Name” and what was written there), structured extraction from complex tables including multi-page ones, and recognition of Indic handwriting. Weaknesses of the first version, mainly hallucinations (a model writing text that is not on the page) and inconsistent output, were also worked on, the post says.
The model does not read a page in one pass. Sarvam describes a “harness-with-VLM” design: a layout parser first works out where the text blocks, tables and figures sit, a reading-order network decides the sequence, and the vision-language model (VLM), a model that takes an image and writes text, reads each piece. According to the post, the VLM can work on a whole page directly, but that accuracy is better with the harness around it.
For the new abilities Sarvam built training sets of handwritten and printed forms, some synthetic, some taken from forms found on the web and then filled in by computer. For handwriting it drew on videos and other sources that contain handwritten text. Training ended with supervised fine-tuning followed by reinforcement learning with verifiable rewards, meaning rewards that can be checked automatically. Angel One’s report of 25 September 2026 repeats these points from the release.
On cost, Sarvam says changes to its inference stack, the software that runs the model for users, let it serve Vision 2.1 at a lower price than it announced in February, and that the API is now tuned for production workloads. No figure is given for the cut. Sarvam’s pricing page, opened on 12 November 2026, lists its Document Digitization API at 0.50 rupee a page and its Document Extraction API at 1.00 rupee a page, excluding GST. That page does not say which model version its prices cover.
The post does not state the model’s size. The February launch post described Sarvam Vision as a 3-billion-parameter state-space vision-language model (parameters are the adjustable numbers inside a model, a rough measure of its size), and Sarvam’s documentation page for the model still says 3 billion. Whether 2.1 keeps that size is not stated.
Why Indian scripts are harder to read
Most Indian scripts descend from Brahmi, and the Unicode Standard’s chapter on them says they share a large number of structural features. Each consonant letter carries an inherent vowel, usually a short “a”, so the letter for “k” is read “ka”. Another vowel is written with a dependent vowel sign, a matra, which can sit above, below, to the left or to the right of the consonant, and in some scripts is built from two parts that wrap around it.
A sign called the virama cancels the inherent vowel, and consonants stripped that way fuse with the next one into a conjunct. Unicode’s own example is the Devanagari cluster K.SSA, which it calls fully ligated, meaning the components are no longer distinct shapes. A well-made Indic font can hold hundreds of conjunct glyphs. Unicode also notes that a few signs change the order in which characters are displayed relative to the order in which they are stored, which is the spoken order, so a program reading pixels has to restore that sequence.
Handwriting adds variation on top. A December 2025 arXiv paper on handwritten Marathi court documents calls handwriting recognition hard for languages like Marathi because large digitised collections are scarce and handwriting styles vary widely. Printed archives bring their own trouble: Sarvam’s Indic test draws on pages from 1800 to the present, and a sample in the dataset viewer shows Bengali textbook sentences with Arabic numerals mixed in.
Three tests, three different questions
olmOCR-Bench comes from the olmOCR paper (February 2025), published with a model for turning PDFs into text. Its dataset card on Hugging Face, hosted by Ai2, lists 1,403 PDFs and 7,010 pass-or-fail checks: is this sentence present, is that page number absent, does the headline come before the paragraph, does a table cell hold the right value, is a formula right. Its pages are in English and include arXiv maths, old scans, tiny print, multi-column pages and tables. Sarvam says the benchmark is officially English-only but contains a few samples in Chinese and other languages. In February it reported on a filtered English-only set (the launch post cites 1,258 of the 1,403 samples). For 2.1 it reports on the full official set, for parity with other makers’ models.
OmniDocBench comes from a December 2024 paper accepted at CVPR 2025. The test spans nine kinds of document source, including academic papers, textbooks, handwritten notes and dense newspapers, and scores a parser on 19 layout categories and 15 attributes. Sarvam reports on version 1.6, which it describes as a combined measure of text errors, table structure and formula recognition, so it tests how faithfully a page’s structure survives.
Sarvam Indic OCR Bench is Sarvam’s own. The Hugging Face card says it holds 6,909 samples of text blocks, cropped from pages, that human language experts reviewed twice. Sarvam’s post counts 6,609 samples across the 22 languages and 300 in English, drawn from newspapers, brochures, textbooks and historical writings. Only language accuracy is measured, scored as word accuracy, which is 100 times one minus the word error rate. The dataset is public under the Apache 2.0 licence, and the card says a scorer using only Python’s standard library is included.
What the scores show
On olmOCR-Bench, Sarvam’s table gives Vision 2.1 an overall 87.3, ahead of Infinity-Parser2 Pro at 86.1 and Opus 5 at 85.1. Vision 2.1 is not first in every column: Infinity-Parser2 Pro scores higher on old scans (58.0 against 55.3), and Opus 5 on multi-column pages (85.8 against 82.1).
On OmniDocBench v1.6, Vision 2.1 scores 94.97 overall. PaddleOCR-VL 1.6 is ahead at 96.01 in Sarvam’s table, and GLM-OCR is behind at 94.71. Vision 2.1 has the lowest text error (0.0289) and the top formula score (0.988) in the table and trails PaddleOCR-VL on table structure (0.890 against 0.931 on the TEDS table measure).
On its Indic bench Sarvam reports 87.39 overall, against 84.94 for Bodhan Indic-OCR, 79.35 for Gemini 3.6 Flash and 71.76 for Google Cloud Vision. Per language, Konkani (97.41), Nepali (97.00) and Maithili (96.70) score highest, with Marathi at 95.06 and Hindi and Bengali at 93.52 and 93.47. Kashmiri (54.82) and Santhali (53.91) score lowest. On Santhali, Bodhan Indic-OCR scores 68.30 in the same table, and Gemini 3.6 Flash leads on Odia, 81.01 against 80.01. The post lists printed sources for this test and no separate handwriting score, so the handwriting ability appears there only in demonstration clips.
What is still to be shown
Sarvam itself says the two English benchmarks are arguably saturated, and that how far a benchmark result predicts real-world usefulness is an open question. Internal tests on real workflows were run too, the post says, and their results are not published.
One check can be made today. The leaderboard on Ai2’s olmOCR-bench page, as opened on 12 November 2026, lists Infinity-Parser2 Pro at 87.6 and another model at 87.84, both above Sarvam’s 87.3, while Sarvam’s own table puts Infinity-Parser2 Pro at 86.1. A company’s comparison table and a public leaderboard can give different numbers for the same model.
The Indic benchmark is the test built around the 22 languages, and it is open for others to run. The wider programme Sarvam works within is covered in our guide to the IndiaAI Mission, the language list in our explainer on the Eighth Schedule, and the government translation platform in our piece on Bhashini and AI4Bharat. The dataset card says a scorer that needs only Python’s standard library comes with the data, so a lab other than Sarvam can publish its own result.
Sources & further reading
- Sarvam AI: Sarvam Vision 2.1, Pushing the Pareto frontier of document intelligence (24 September 2026)
- Sarvam AI: Sarvam Vision, launch post (5 February 2026)
- Sarvam AI: blog index with post dates (opened 12 November 2026)
- Sarvam API docs: Sarvam Vision model page (languages, size, limits)
- Sarvam API docs: Pricing, credits and rate limits (opened 12 November 2026)
- Hugging Face: Sarvam Indic OCR Bench, dataset card
- Hugging Face: olmOCR-bench, dataset card and leaderboard (Ai2)
- arXiv: olmOCR, Unlocking Trillions of Tokens in PDFs with Vision Language Models (February 2025)
- arXiv: OmniDocBench, Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations (December 2024)
- The Unicode Standard 15.0, chapter 12: South and Central Asia-I (Devanagari and the Indic script principles)
- arXiv: Seeing Justice Clearly, Handwritten Legal Document Translation with OCR and Vision-Language Models (December 2025)
- Analytics India Magazine: Sarvam Vision Beats Gemini, GPT on Indic OCR Across 22 Indian Languages (5 February 2026)
- Angel One News: Sarvam Launches Vision 2.1 AI Model for Documents and Indian Languages (25 September 2026)
Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.
Enjoyed this? Get the next one.
One good read at a time, straight to your inbox. No spam, unsubscribe anytime.