- Math500 98.6
- Live Code Bench v6 71.7
- MMLU 90.6
- MMLU Pro 81.7
- Writing Bench (Hugging Face card table, official critic model) 80.5
- Arena Hard v2 71
- IF Eval 84.8
- GPQA Diamond 78.7
- AIME 25 (88.3 without tools, 96.7 with tools) 88.3
- HMMT (Feb 25) 85.8
- HMMT (Nov 25) 85.8
- Beyond AIME 69.1
- BrowseComp 49.5
- SWE Bench Verified (SWE-Agent Harness) 45
- Tau2 (avg.) 68.3
- HLE 11.2
- Average pairwise win rate across all dimensions on the Indic benchmark (indivibe) 90
Sarvam 105B
Sarvam 105B places in the top three on MATH-500.
16
Benchmarks with a score
0
Ranked first, out of five or more models
1
In the top three
Sarvam AI Mar 2026
Sarvam 105B
Open weights Flagship
No Epoch Capabilities Index score yet
- BrowseComp 49.5% #59 /77
- tau2-bench 68.3% #54 /69
- MMLU-Pro 81.7% #60 /151
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMLU | 90.6% | #9 of 127 | Model card | |
| MMLU-Pro | 81.7% | #60 of 151 | Model card | |
| Arena-Hard v2.0 | 71% | #11 of 24 | Model card | |
| IndiVibe | 90% | #1 of 2 | Model card | |
| IFEval | 84.8% | #77 of 126 | Model card | |
| Humanity's Last Exam | 11.2% | #128 of 147 | Model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH-500 | 98.6% | #2 of 67 | Model card | |
| AIME 2025 | 88.3% | #59 of 143 | Model card | |
| GPQA Diamond | 78.7% | #119 of 284 | Model card | |
| Beyond AIME | 69.1% | #6 of 14 | Model card | |
| HMMT 2025 | 85.8% | #28 of 49 | Model card | |
| HMMT Nov 2025 | 85.8% | #18 of 23 | Model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench v6 | 71.7% | #45 of 76 | Model card | |
| SWE-bench Verified | 45% | #132 of 150 | Model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BrowseComp | 49.5% | #59 of 77 | Model card | |
| tau2-bench | 68.3% | #54 of 69 | Model card |
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.