Sarvam-30B
Sarvam-30B is an open-source 30B-parameter Mixture-of-Experts reasoning model from Sarvam AI trained from scratch and optimized for Indian languages, coding, and conversational workloads. It uses 128 sparse experts with 2.4B active parameters per token, Grouped Query Attention, and was pre-trained on 16 trillion tokens spanning code, mathematics, multilingual, and web data.
Benchmark results
| Benchmark | Score | Tags | Source |
|---|---|---|---|
| AIME 2025 | 96.7% | self-reported llm-stats | link → |
| Arena-Hard v2 | 49.0% | self-reported llm-stats | link → |
| Beyond AIME | 58.3% | self-reported llm-stats | link → |
| BrowseComp | 35.5% | self-reported llm-stats | link → |
| GPQA | 66.5% | self-reported llm-stats | link → |
| HMMT 2025 | 73.3% | self-reported llm-stats | link → |
| HMMT25 | 74.2% | self-reported llm-stats | link → |
| HumanEval | 92.1% | self-reported llm-stats | link → |
| LiveCodeBench v6 | 70.0% | self-reported llm-stats | link → |
| MATH-500 | 97.0% | self-reported llm-stats | link → |
| MBPP | 92.7% | self-reported llm-stats | link → |
| MMLU | 85.1% | self-reported llm-stats | link → |
| MMLU-Pro | 80.0% | self-reported llm-stats | link → |
| SWE-Bench Verified | 34.0% | self-reported llm-stats | link → |