Sarvam Vision 2.1 launches with SOTA OCR scores and Indic handwriting recognition across 22 languages
itsOmSarraf_ · x · 2026-09-24
Indian AI firm Sarvam AI released Sarvam Vision 2.1, a document intelligence model adding structured extraction and Indic handwriting recognition across English and 22 Indian languages.
- State of the art on both olmOCR-Bench (87.3) and its own Indic benchmark (87.39)
- New capabilities for complex multi-page table parsing and key-value extraction from forms
- Inference stack optimizations enable cheaper pricing than at launch, with the API tuned for production workloads
- Hallucination and inconsistency issues addressed
Part of its sovereign models lineup launched in February 2026, built on the harness-with-VLM architecture (semantic layout parser plus pointer reading order).
Related event: Sarvam Vision 2.1 Tops OCR Benchmarks Across 22 Indian Languages(2 posts)→
More from Models
- Boltzbit's BAST paper claims LLMs can learn 1000x faster by generating weights from live data — ahuja_priyank · 2026-09-24
- Fireworks launches Ember-1: Kimi K3-based model cuts reasoning tokens by ~40% — sophiamyang · 2026-09-24
- Liquid AI's DSpark speculative decoding makes LFM2.5-VL-3B up to 3.13x faster — helloiamleonie · 2026-09-24
- Users can't get ChatGPT to brighten a photo without regenerating faces — Objective-Bit-797 · 2026-09-24
- Opus 5.5 builds a 40-second animation and explains how it did it — zck · 2026-09-24
- DeepSeek Harness Quietly Ships Official GUI Client for Windows and Mac — vista8 · 2026-09-24