Sarvam 105B: benchmark results
Sarvam AI MoE model with 10.3B active parameters, trained from scratch in India. The publisher reports 12T pretraining tokens and 128K context with YaRN scaling. Provider: Sarvam AI. Released 2026-02-18. Access: Open.
Aggregate Rating 1609 ± 30, rank #409 of 1656 rated models, from 62 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Beyond AIME | 69.1 | Score (percent) | 100 |
| ZeroEval MATH-500 | 98.6 | MATH-500 Score | 96.9 |
| Enkrypt AI - Insecure Code Risk | 6.67 | Risk Score | 87.7 |
| Enkrypt AI - Risk Score | 22.53 | Risk Score | 81.8 |
| Enkrypt AI - Toxicity Risk | 1.64 | Risk Score | 77.4 |
| Enkrypt AI - Jailbreak Risk | 6.47 | Risk Score | 69 |
| AA Omniscience - Software Engineering (SWE) - R | 14 | Accuracy (%) | 67.6 |
| AA Omniscience - Software Engineering (SWE) - Julia | 12 | Accuracy (%) | 58 |
| ZeroEval GPQA Diamond | 78.7 | GPQA Diamond Score | 57.2 |
| Arcee MMLU-Pro (BenchLM) | 81.7 | Accuracy (%) | 54.8 |
| AA GPQA Diamond | 73.84 | Accuracy (%) | 54.7 |
| LLM Stats Score | 25.41 | LLM Stats Score (conservative rating) | 54.3 |
Interactive version: theaggregate.ai/model?slug=sarvam-105b · How It Works · Data refreshed daily, snapshot 2026-10-20.