Sarvam AI Drops Sarvam-105B: India's Open-Source AI Powerhouse That Rivals DeepSeek and Frontier Models

TamilTech AI summary

Sarvam AI, the Bengaluru-based lab picked by the Indian government for sovereign AI work, has open-sourced Sarvam-105B, a 105-billion-parameter Mixture-of-Experts model that only activates about 10.3B parameters per token, so you get big-model reasoning at a much lower compute cost. It ships under Apache 2.0 alongside a smaller Sarvam-30B option, supports a 128K context window, and is built for English plus 22 Indian languages, which is a real first at this scale. On benchmarks it holds its own or beats strong peers like GLM-4.5-Air, GPT-OSS-120B, and Qwen3-80B on math, coding, MMLU, and especially agentic tasks such as BrowseComp, and it even edges DeepSeek-class results on several reasoning tests despite far fewer active parameters. That matters because India is no longer only consuming Silicon Valley or China-built AI—it now has a commercially usable, fully open frontier-style model trained with Indic languages in mind for products, enterprises, and public services. If you want to try it, the weights are on Hugging Face at sarvamai/sarvam-105b, with API access and vLLM/SGLang deployment notes so developers can start building right away.

• What is Sarvam-105B?
• Does Sarvam-105B beat DeepSeek?
• What languages does Sarvam-105B support?
• How can I use Sarvam-105B?

AI-assisted summary, checked by the TamilTech editorial team.