Sarvam open-sources 30B, 105B models

This is the kind of release that matters more than a flashy leaderboard screenshot — it shows a regional lab trying to build real sovereign AI infrastructure, not just wrap someone else’s model.

• –Sarvam says both models were trained from scratch in-house, which is a stronger signal than a fine-tuned derivative release
• –The 105B model is positioned as globally competitive on reasoning, programming, and agentic benchmarks, including strong tool-use results
• –The release leans hard into Indian-language performance and tokenizer efficiency, which could make it more practical than generic open models for India-focused products
• –Open weights plus API access give developers two paths: self-host for control or use Sarvam’s stack for faster deployment
• –The real test now is community adoption: if developers validate the coding and agentic claims outside Sarvam’s own demos, this could become a meaningful open-model alternative