Quick answer: Sarvam 105B is Sarvam AI's July 2025 large model scoring 71.0% Arena Hard v2 and 81.7% MMLU-Pro. Apache 2.0 — India-focused AI lab's largest model.
Where Sarvam 105B leads
Where it lags
Best for: Indian-language + English mixed deployments; Apache 2.0 105B scale; Indic language tasks.
Sarvam 105B is the flagship model from Sarvam AI (India), released July 2025. The Indian AI startup focuses on multilingual models supporting Indic languages alongside English. The 81.7% MMLU-Pro is competitive for a July 2025 105B model.
| Field | Value |
|---|---|
| Organization | Sarvam AI |
| License | Apache 2.0 |
| HuggingFace | sarvamai/sarvam-105b |
| Release date | July 2025 |
| Parameters | 105B |
| Modality | Text only |
Open weights under Apache 2.0 — self-host at no cost.
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Arena Hard v2 | 71.0% | Benchgen evaluation | 2025-07 |
| MMLU-Pro | 81.7% | Benchgen evaluation | 2025-07 |
| Model | Arena Hard v2 | MMLU-Pro | License |
|---|---|---|---|
| Sarvam 105B | 71.0% | 81.7% | Apache 2.0 |
| Sarvam 30B | 53.0% | 72.4% | Apache 2.0 |
| Qwen2.5 72B Instruct | — | 79.8% | Apache 2.0 |
Sarvam 105B vs 30B: +18% Arena Hard (71% vs 53%), +9.3% MMLU-Pro. Worth the scale-up if 105B budget is available.
Specs from Sarvam AI's Sarvam 105B release (July 2025) and Benchgen evaluations. Last updated 2026-07-24.
This model isn’t on any benchmark leaderboard yet.