Open-Sourcing Sarvam 30B and 105B
Summary
The article announces open-sourcing Sarvam 30B and 105B, detailing architecture (Mixture-of-Experts Transformer with GQA and MLA), end-to-end in-house training, and inference optimization across hardware. It covers benchmarks, Indian language performance, tokenizer efficiency, and live demos, underscoring India's sovereign AI stack and deployment readiness.