Sarvam 30B: benchmark results
Sarvam AI MoE model trained from scratch in India on 16T tokens. Its publisher reports 2.4B non-embedding active parameters; that count excludes embeddings. The model targets multilingual conversation, reasoning, and tool use. Provider: Sarvam AI. Released 2026-02-18. Access: Open.
Aggregate Rating 1520 ± 25, rank #719 of 1656 rated models, from 72 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ZeroEval MATH-500 | 97 | MATH-500 Score | 76.6 |
| Enkrypt AI - Insecure Code Risk | 20.89 | Risk Score | 59.8 |
| Enkrypt AI - Risk Score | 29.31 | Risk Score | 59.5 |
| Enkrypt AI - Toxicity Risk | 3.27 | Risk Score | 51.2 |
| Arcee MMLU-Pro (BenchLM) | 80 | Accuracy (%) | 50.7 |
| AA CritPt | 0.29 | Accuracy (%) | 48.2 |
| AA Omniscience - Software Engineering (SWE) - Julia | 8 | Accuracy (%) | 46.1 |
| Tau2-Bench Telecom | 34.5 | Success Rate (%) | 43.6 |
| AA Humanity's Last Exam | 7.51 | Accuracy (%) | 42.9 |
| AA TAU-2 Bench | 34.5 | Accuracy (%) | 42.5 |
| LLM Stats Score | 19.22 | LLM Stats Score (conservative rating) | 42.3 |
| AA Omniscience - Software Engineering (SWE) - Go | 14 | Accuracy (%) | 41.1 |
Interactive version: theaggregate.ai/model?slug=sarvam-30b · How It Works · Data refreshed daily, snapshot 2026-10-20.