Indian startup Sarvam launches AI models to rival global giants
Indian startup Sarvam just launched two homegrown AI models—30-billion and 105-billion parameter models—at the India AI Impact Summit, aiming to take on global tech giants.
Both models work in all 22 Indian languages, support voice-first chats, and use a clever architecture that keeps things fast and efficient.
The 2 models are designed for different use cases
The 30-billion model packs 30 billion parameters and can handle long conversations (up to 32,000 tokens). It powers conversational products and supports code-mixed and Hinglish input.
The 105-billion parameter model is even bigger—105 billion parameters with a massive 128,000-token window—and posts strong benchmark results against other large models on OCR and Indic-language tasks. It supports document understanding and long-context tasks.
Vision model outperforms Gemini and ChatGPT on Indic-language tasks
Sarvam's Vision model tops Gemini and ChatGPT at reading Indian scripts (over 84% accuracy in Indic OCR). The company plans to open-source both models.
More tools like Samvaad agents and Sarvam for Work are coming up.
Backed by over $50 million from investors including Lightspeed and Khosla Ventures, plus support from IndiaAI Mission, Sarvam is putting Indian languages front-and-center in the global AI race.