Sarvam releases 30B and 105B model weights under Apache 2.0

Sarvam AI/Sarvam 30B / 105B

Sarvam published the weights of both models on Hugging Face and the government's AIKosh platform under the permissive Apache 2.0 licence. The 30B model runs Sarvam's Samvaad conversational platform and the 105B model its Indus assistant.

• 30B activates about 2.4B parameters per token; 105B uses latent attention for long inputs.
• 105B reported 98.6 on Math500 and 90.6 on MMLU.
• Training covered pre-training, supervised fine-tuning and reinforcement learning, all in India.