Sarvam AI Unveils India-Trained Open-Source LLMs: Sarvam 30B & 105B Models Released

• Sarvam AI open-sourced India-trained Sarvam 30B and Sarvam 105B large language models, advancing domestic AI infrastructure.
• Models built from scratch using IndiaAI mission compute resources, covering pre-training, fine-tuning, and reinforcement learning.
• Zoho co-founder Sridhar Vembu lauded the release, stressing the importance of sustained R&D for innovation.
• Utilize Mixture-of-Experts Transformer architecture; Sarvam 30B powers Samvaad, Sarvam 105B powers Indus AI assistant.
• Available via Sarvam AI API, AI Kosh, and Hugging Face under Apache 2.0 open-source license.