Last updated:
Sarvam
Sarvam AI's open models trained from scratch in India for the country's 22 official languages.
Model Facts
Latest model: Sarvam-30B
- Provider
- Sarvam AI
- Released
- Feb 19, 2026
- Weights
- Open
Full Model List
| Model | Released | Status | Note |
|---|---|---|---|
| Sarvam-30B | Feb 19, 2026 | current | 30B model trained from scratch in India |
| Sarvam-105B | Feb 19, 2026 | current | 105B MoE (10.3B active), trained from scratch |
| Sarvam-M | May 1, 2025 | legacy | fine-tuned from Mistral Small; month approximate |
- Official model page:
- sarvam.ai
- Official documentation:
- docs.sarvam.ai
- Other sources:
- Sarvam AI (Wikipedia)
- Sarvam-105B (Artificial Analysis)
What it is
Sarvam AI announced Sarvam-30B and Sarvam-105B in February 2026 and released them as open weights under Apache 2.0 (Hugging Face and AIKosh) in early March. Both were trained from scratch in India on the IndiaAI Mission's compute, with a focus on all 22 Indian languages, reasoning, mathematics and coding; Sarvam-105B is a mixture-of-experts model with 10.3B active parameters. They succeed Sarvam-M (May 2025), which was fine-tuned from Mistral Small.
Where this information comes from
Details come from Indian press coverage and an Artificial Analysis write-up. Context windows were not confirmed. Release dates, context windows and benchmark scores are vendor-reported unless stated, and harnesses differ between labs, so treat the numbers as indicative.
Notable traits
- Voice-first optimisation for Indian languages.
- Apache 2.0 weights.
- Trained from scratch rather than fine-tuned.