Large language model
Verdict
Sarvam's smaller India-trained MoE (30B / 2.4B active). Powers Samvaad. Apache 2.0, 32K context. The laptop-to-single-GPU Indic option.
“Practical Indic workhorse if 105B is too big.”
Start here for Indic chat; step up to 105B for harder reasoning.
Training data policy: N/A — open weights. Self-host for full control.
DPDP Act (India): Trained in India. Self-host or use India-region API for DPDP control.
Data residency: India, Self-hosted
Self-host in India