Sarvam AI Draws Global Attention After Sundar Pichai’s Praise at India AI Impact Summit 2026

Sarvam AI, a Bengaluru-based artificial intelligence startup focused on building models for Indian languages and local use cases, came into focus at the India AI Impact Summit 2026 after Google CEO Sundar Pichai highlighted its work on homegrown AI systems.

Speaking at the summit, Pichai referred to Sarvam AI's efforts in developing local AI models suited to India's linguistic needs. "The work Sarvam has done developing local AI models...I just don't see any impediments to that, and I think it is very, very well positioned," he said.

The remarks come as Sarvam AI continues to position its in-house models as alternatives built specifically for multilingual environments common across India.

Sarvam Vision Model Benchmark Claims

Sarvam AI stated that its Sarvam Vision model recorded 84.3% accuracy on the olmOCR-Bench (English-only subset), outperforming frontier models such as Gemini 3 Pro and DeepSeek OCR 2, according to CEO Pratyush Kumar.

The model supports image captioning, scene text recognition, chart reading, and complex table parsing across varied layouts. The company says these capabilities are intended for analysing scanned files, dense reports, and mixed-format records containing text, figures, and graphical elements.

Focus on Indian Languages

Sarvam AI's vision-language model has been trained on datasets covering all 22 official Indian languages, including financial paperwork, newspapers, literature, historical archives, and public documents.

The company has argued that many global AI systems still perform best in English and treat regional scripts as secondary inputs. Its stated aim is to improve multilingual document understanding and enable knowledge extraction from scanned or legacy records that are not yet digitised.

On-Device Speech and Translation Models

Sarvam AI has also developed speech recognition, speech synthesis, and translation systems designed for on-device deployment:

Speech recognition model: 74 million parameters (~294MB), supports 10 Indian languages with automatic language detection

Speech synthesis model: 24 million parameters (~60MB), supports custom voice cloning using about one hour of recorded audio

Translation model: 150 million parameters (~334MB), supports bidirectional translation across 110 language pairs involving English and Indian languages

The speech recognition system reportedly operates at around 8.5 times real-time speed, with a time-to-first-token below 300 milliseconds on a Snapdragon 8 Gen 3 chipset.

Document Intelligence API

Sarvam AI has opened access to its Document Intelligence API based on the Sarvam Vision model. The company said the API will remain free through February 2026 for developers testing large-scale document processing applications.

With products spanning vision, speech, and translation, Sarvam AI is focusing on multilingual document understanding and India-specific deployment scenarios as interest grows in locally trained AI systems.