Sarvam-105B
Sarvam-105B is Sarvam AI's flagship open-source Mixture-of-Experts reasoning model built for complex reasoning, coding, and agentic workflows. It uses 128 sparse experts with Multi-head Latent Attention for efficient long-context inference and was pre-trained on 12 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 | 71.0% | self-reported llm-stats | link → |
| Beyond AIME | 69.1% | self-reported llm-stats | link → |
| BrowseComp | 49.5% | self-reported llm-stats | link → |
| GPQA | 78.7% | self-reported llm-stats | link → |
| HMMT 2025 | 85.8% | self-reported llm-stats | link → |
| HMMT25 | 85.8% | self-reported llm-stats | link → |
| Humanity's Last Exam | 11.2% | self-reported llm-stats | link → |
| IFEval | 84.8% | self-reported llm-stats | link → |
| LiveCodeBench v6 | 71.7% | self-reported llm-stats | link → |
| MATH-500 | 98.6% | self-reported llm-stats | link → |
| MMLU | 90.6% | self-reported llm-stats | link → |
| MMLU-Pro | 81.7% | self-reported llm-stats | link → |
| SWE-Bench Verified | 45.0% | self-reported llm-stats | link → |