Sarvam unveils 30B and 105B mixture-of-experts models trained from scratch in India

Sarvam AI/Sarvam 30B / 105B

At the India AI Impact Summit in New Delhi, Sarvam showed two new language models built on IndiaAI Mission compute with help from Yotta and Nvidia. The 30B model targets real-time conversation and the 105B model multi-step reasoning; Sarvam said both would be open-sourced.

• 30B: 32K-token context, about 16 trillion training tokens; pitched against Gemma 27B and GPT-OSS-20B.
• 105B: 128K-token context; pitched against GPT-OSS-120B and Qwen3-Next-80B.
• Launched alongside speech-to-text, text-to-speech and document-vision models.