The company says it adds stronger multi-page table parsing, key-value extraction from forms, and Indic handwritten-text recognition. Its architecture combines a VLM with a semantic layout parser and a reading-order network.
Sarvam reports an overall score of 87.30 on olmOCR-Bench, versus 86.10 for Infinity-Parser2 Pro and 85.10 for Opus 5 in its comparison table. On OmniDocBench v1.6, it reports 94.97, behind PaddleOCR-VL 1.6 at 96.01. The post also says inference-stack optimization lowered the serving price from the level announced at the original launch, without stating the new price.
The company describes the release as optimized for production workloads and says it addressed hallucinations and inconsistencies. The source does not provide independent benchmark validation, a public price, or evidence of performance on a specific customer's documents.