Indian AI startup Sarvam announced at its Epoch 2026 conference that it is building a foundational model with more than one trillion parameters from scratch in India. Co-founder Pratyush Kumar said the model aims to compete in coding, cybersecurity, simulation and science. The company also upgraded its flagship Sarvam 105B model, claiming stronger results on voice calls, instruction following and agent tasks than OpenAI's GPT-5.4 Mini and Google's Gemini 3.5 Flash. Sarvam says 105B costs $0.80 per million blended tokens, making it 5.5 times cheaper than GPT-5.4 Mini and 11 times cheaper than Gemini 3.5 Flash. It launched Vision 2.0 with improved OCR for Indian handwritten text, tables and key-value extraction, plus commercial Vision Edge for on-device document work. Speech models Saras V4 and Bulbul V4 add better recognition of languages such as Odia, Sanskrit and Manipuri, plus emotional text-to-speech. A demo of Kaze smartglasses for the visually impaired was also shown.
Indian AI startup Sarvam announced at its Epoch 2026 conference that it is building a foundational model with more than one trillion parameters from scratch in India. Co-founder Pratyush Kumar said the model aims to compete in coding, cybersecurity, simulation and science. The company also upgraded its flagship Sarvam 105B model, claiming stronger results on voice calls, instruction following and agent tasks than OpenAI's GPT-5.4 Mini and Google's Gemini 3.5 Flash. Sarvam says 105B costs $0.80 per million blended tokens, making it 5.5 times cheaper than GPT-5.4 Mini and 11 times cheaper than Gemini 3.5 Flash. It launched Vision 2.0 with improved OCR for Indian handwritten text, tables and key-value extraction, plus commercial Vision Edge for on-device document work. Speech models Saras V4 and Bulbul V4 add better recognition of languages such as Odia, Sanskrit and Manipuri, plus emotional text-to-speech. A demo of Kaze smartglasses for the visually impaired was also shown.