{"id":7069484,"date":"2026-09-29T00:52:08","date_gmt":"2026-09-29T00:52:08","guid":{"rendered":"https:\/\/peraltafinancing.com\/analytics\/indic-ocr-benchmarks-and-features\/"},"modified":"2026-09-29T00:52:08","modified_gmt":"2026-09-29T00:52:08","slug":"indic-ocr-benchmarks-and-features","status":"publish","type":"post","link":"https:\/\/fivemor.com\/?p=7069484","title":{"rendered":"Indic OCR Benchmarks and Features"},"content":{"rendered":"


\n<\/p>\n

\n

For the first time, an OCR model reads Indian languages as fluently as English documents. Sarvam Vision 2.1 breaks a long-standing trade-off. Indian businesses had to choose between structural document parsing and script recognition. Never both. Finance teams automating invoices faced this. Insurance companies processing handwritten claims in multiple states faced this. Organizations digitizing regional records faced this. All had to pick accuracy or coverage. This model does both. Let\u2019s explore what changed, where it wins, where it struggles, and five test documents you can run yourself.<\/p>\n

Key Features of Sarvam Vision 2.1<\/h2>\n

Pulling named fields out of a table rather than transcribing the entire grid. This works for statements, ledgers, or anything where a value only makes sense in relation to its row and column headers.<\/p>\n

The same idea applied to form fields. The label and value sit in separate boxes. The pairing has to be inferred from layout, not reading order.<\/p>\n

Indic Handwritten Extraction<\/h3>\n

Handwriting in Indian scripts extracted into structured fields rather than transcribed as a block. This is the hardest capability. Sarvam trained it partly on video sources to capture real handwriting variety, not just synthetic samples.<\/p>\n

What the Numbers Actually Show<\/h2>\n

Until now, if a model was good at English document parsing it was usually mediocre at Indian languages, and the models that handled Indic scripts well were not competitive on general document structure. Those were separate tools with separate failure modes.<\/p>\n

\"Sarvam<\/figure>\n

Each point is one system scored on both axes. Upper right is better on both at once.<\/p>\n

Look at where the other points sit. Infinity-Parser2 Pro scores 86.1 on English, second only to Sarvam, and then collapses to 49.83 on Indic. Google Cloud Vision does the reverse: 39.6 on English document structure, but 71.76 on Indian languages, because it has had Indic OCR for years without the layout intelligence. Sarvam 2.1 is the only point in the upper right.<\/p>\n

Where it Wins and Where it Doesn\u2019t<\/h2>\n

The Indic Benchmark<\/h3>\n

Sarvam also released the benchmark itself: 6,909 samples, 6,609 spanning all 22 official Indian languages and 300 in English, drawn from newspapers, brochures,\u00a0textbooks\u00a0and historical writing dated from 1800 to the present.<\/p>\n

The benchmark is\u00a0published on Hugging Face<\/a>, which matters. A vendor-built benchmark that the vendor wins\u00a0is\u00a0worth little on its own. A vendor-built benchmark released publicly so others can run it is a different proposition, and it is the right way to do this.<\/p>\n

olmOCR-Bench<\/h2>\n

The community benchmark for English document parsing. It runs pass-fail checks across eight categories:\u00a0arXiv\u00a0maths, base text, headers and footers, tiny text, multi-column pages, degraded old scans, old-scan maths, and tables. It tests whether facts arepresent or absent rather than scoring subtle differences.<\/p>\n

Sarvam notes that this benchmark is officially English-only but\u00a0contains\u00a0some contaminant samples in Chinese and other scripts. Their\u00a0previous\u00a0release reported on a filtered English-only set; this time they report on the official set for parity with competitors, which is the more conservative choice.\u00a0<\/p>\n

\n\n\n\n\n\n\n\n\n\n\n
Model<\/strong>\u00a0<\/td>\nMath<\/strong>\u00a0<\/td>\nTables<\/strong>\u00a0<\/td>\nOldScan<\/strong>\u00a0<\/td>\nMultCol<\/strong>\u00a0<\/td>\nOverall<\/strong>\u00a0<\/td>\n<\/tr>\n
Sarvam Vision 2.1\u00a0<\/td>\n90.5\u00a0<\/td>\n91.9\u00a0<\/td>\n55.3\u00a0<\/td>\n82.1\u00a0<\/td>\n87.3\u00a0<\/td>\n<\/tr>\n
Infinity-Parser2 Pro\u00a0<\/td>\n87.4\u00a0<\/td>\n88.9\u00a0<\/td>\n58.0\u00a0<\/td>\n83.3\u00a0<\/td>\n86.1\u00a0<\/td>\n<\/tr>\n
Opus 5\u00a0<\/td>\n90.0\u00a0<\/td>\n89.5\u00a0<\/td>\n54.0\u00a0<\/td>\n85.8\u00a0<\/td>\n85.1\u00a0<\/td>\n<\/tr>\n
Chandra-OCR2\u00a0<\/td>\n86.5\u00a0<\/td>\n87.5\u00a0<\/td>\n49.2\u00a0<\/td>\n82.4\u00a0<\/td>\n84.5\u00a0<\/td>\n<\/tr>\n
Mistral OCR4\u00a0<\/td>\n83.7\u00a0<\/td>\n88.6\u00a0<\/td>\n48.9\u00a0<\/td>\n85.7\u00a0<\/td>\n83.1\u00a0<\/td>\n<\/tr>\n
Gemini 3.6 Flash\u00a0<\/td>\n86.5\u00a0<\/td>\n85.9\u00a0<\/td>\n48.1\u00a0<\/td>\n78.6\u00a0<\/td>\n82.4\u00a0<\/td>\n<\/tr>\n
GPT 6 Astra\u00a0<\/td>\n82.6\u00a0<\/td>\n90.9\u00a0<\/td>\n47.0\u00a0<\/td>\n77.8\u00a0<\/td>\n81.8\u00a0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

OmniDocBench\u00a0v1.6<\/strong>\u00a0<\/h2>\n

A different measure: structural fidelity rather than fact presence. It is a composite of text edit distance, table structure scored with TEDS, formula recognition scored with CDM, and reading order, run over newspapers, textbooks,\u00a0magazines\u00a0and financial reports.<\/p>\n

\n\n\n\n\n\n\n\n\n\n
Model<\/strong>\u00a0<\/td>\nText edit\u00a0dist\u00a0(lower better)<\/strong>\u00a0<\/td>\nFormula CDM<\/strong>\u00a0<\/td>\nTable TEDS<\/strong>\u00a0<\/td>\nOverall<\/strong>\u00a0<\/td>\n<\/tr>\n
PaddleOCR-VL 1.6\u00a0<\/td>\n0.0356\u00a0<\/td>\n0.985\u00a0<\/td>\n0.931\u00a0<\/td>\n96.01\u00a0<\/td>\n<\/tr>\n
Sarvam Vision 2.1\u00a0<\/td>\n0.0289\u00a0<\/td>\n0.988\u00a0<\/td>\n0.890\u00a0<\/td>\n94.97\u00a0<\/td>\n<\/tr>\n
GLM-OCR\u00a0<\/td>\n0.0374\u00a0<\/td>\n0.984\u00a0<\/td>\n0.895\u00a0<\/td>\n94.71\u00a0<\/td>\n<\/tr>\n
GPT 6 Astra\u00a0<\/td>\n0.0460\u00a0<\/td>\n0.967\u00a0<\/td>\n0.891\u00a0<\/td>\n93.74\u00a0<\/td>\n<\/tr>\n
Gemini 3.6 Flash\u00a0<\/td>\n0.0371\u00a0<\/td>\n0.976\u00a0<\/td>\n0.869\u00a0<\/td>\n93.58\u00a0<\/td>\n<\/tr>\n
Opus 5\u00a0<\/td>\n0.0471\u00a0<\/td>\n0.967\u00a0<\/td>\n0.856\u00a0<\/td>\n92.51\u00a0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Read that table across rather than down. Sarvam has the best text edit distance of any model at 0.0289 and the best formula score at 0.988. It loses the top spot purely on table structure, where\u00a0PaddleOCR-VL scores 0.931 against Sarvam\u2019s 0.890. Sarvam calls both benchmarks\u00a0arguably saturated, which is fair when the top twelve models sit inside four points of each other.<\/p>\n

On the Global Benchmarks\u00a0<\/h2>\n

Sarvam leads\u00a0olmOCR-Bench at 87.3 overall. It does not lead every category:\u00a0<\/p>\n

\n\n\n\n\n\n\n\n\n
Category<\/strong>\u00a0<\/td>\nSarvam 2.1<\/strong>\u00a0<\/td>\nBest score<\/strong>\u00a0<\/td>\nHeld by<\/strong>\u00a0<\/td>\n<\/tr>\n
Old scans\u00a0<\/td>\n55.3\u00a0<\/td>\n58.0\u00a0<\/td>\nInfinity-Parser2 Pro\u00a0<\/td>\n<\/tr>\n
Multi-column\u00a0<\/td>\n82.1\u00a0<\/td>\n85.8\u00a0<\/td>\nOpus 5\u00a0<\/td>\n<\/tr>\n
Tiny text\u00a0<\/td>\n92.5\u00a0<\/td>\n93.5\u00a0<\/td>\nOpus 5\u00a0<\/td>\n<\/tr>\n
Tables\u00a0<\/td>\n91.9\u00a0<\/td>\n91.9\u00a0<\/td>\nSarvam 2.1\u00a0<\/td>\n<\/tr>\n
Math\u00a0<\/td>\n90.5\u00a0<\/td>\n90.5\u00a0<\/td>\nSarvam 2.1\u00a0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

And on\u00a0OmniDocBench\u00a0v1.6 Sarvam is second, not first: 94.97 against\u00a0PaddleOCR-VL 1.6 at 96.01. Sarvam wins on text edit distance and formula recognition,\u00a0PaddleOCR\u00a0wins on table structure with a TEDS of 0.931 against 0.890.\u00a0<\/p>\n

Santhali\u00a0is the clear loss. Sarvam scores 53.91 and Bodhan Indic-OCR scores 68.30, a gap of more than fourteen points. Odia is a narrower loss to Gemini 3.6 Flash, 80.01 against 81.01. Kashmiri is not a\u00a0loss\u00a0but it is weak in absolute terms at 54.82, the best score any model manages on that language.\u00a0<\/p>\n

Sarvam Vision 2.1 Architecture\u00a0<\/h2>\n
\"\"<\/figure>\n