The brief

Analyst view

• Sarvam 30B and 105B, trained from scratch on government-funded GPUs, became open weights (Apache 2.0) in March 2026.
• Series B: $234M first close in June 2026 at a $1.5B valuation, $150M of it from HCLTech.
• Next: a model of over 1 trillion parameters for coding, cybersecurity and science, announced in July 2026.

Technical approach

As reported

From scratch

Own models, not fine-tunes

Sarvam pre-trains its own models instead of adapting Western ones, so the data mix and tokenizer can favour Indian languages.

Sparse

Mixture-of-experts to cut cost

Its big models switch on only a small slice of their weights per word, keeping serving cheap enough for India-scale use.

Voice

Speech in and out

Speech-to-text and text-to-speech models wrap the LLMs, because many users would rather talk than type in their own language.

Edge

Small models on devices

Models of a few megabytes, tuned for Qualcomm chips, run offline; HMD is adding an assistant to Nokia feature phones, Bosch is working on cars.

Show all 4