Sarvam AIopen-weight
Released: 2024-08-14Sarvam 2B (Sarvam-1)
Groundbreaking 2-billion parameter foundational language model pre-trained on 4 trillion tokens with extensive native representation for 10 Indian languages and English.
Intelligence Index
76.5/ 100
Calibrated multi-domain compositeThroughput Speed
165tok/sec
Output streaming throughputTime To First Token
140ms
Initial chunk server latencyLiveBench Contam-Free
66.8%
September 2026 suite scoreSourced Benchmark Evaluations (61)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| LiveBench (Continuous) | Active | 66.8% | 94 | 2025-02-01 | independent | Source ↗ |
| ARC-AGI-3 | Active | 48.8% | 93 | 2024-09-18 | independent | Source ↗ |
| ARC-AGI-2 | Nearing Saturation | 58.5% | 75 | 2024-09-11 | independent | Source ↗ |
| FrontierMath | Active | 20.2% | 98 | 2024-09-08 | independent | Source ↗ |
| PutnamBench | Active | 37.4% | 90 | 2024-09-07 | independent | Source ↗ |
| Omni-MATH | Deprecated | 53.3% | 10 | 2024-09-05 | independent | Source ↗ |
| SWE-Lancer | Deprecated | 46.2% | 20 | 2024-09-05 | independent | Source ↗ |
| Soohak | Active | 54.7% | 98 | 2024-09-03 | independent | Source ↗ |
| FrontierCode | Active | 39.4% | 98 | 2024-09-03 | independent | Source ↗ |
| Humanity's Last Exam | Active | 33% | 93 | 2024-09-01 | independent | Source ↗ |
| StructFlowBench | Nearing Saturation | 59.8% | 75 | 2024-09-01 | independent | Source ↗ |
| AIME | Nearing Saturation | 68.4% | 75 | 2024-09-01 | independent | Source ↗ |
| OlympiadBench | Active | 59% | 80 | 2024-08-31 | independent | Source ↗ |
| FireBench | Active | 59.1% | 95 | 2024-08-30 | independent | Source ↗ |
| Terminal-Bench | Active | 47.6% | 80 | 2024-08-30 | independent | Source ↗ |
| MathArena | Active | 57.6% | 99 | 2024-08-30 | independent | Source ↗ |
| OpenAI-MRCR | Active | 64.3% | 80 | 2024-08-29 | independent | Source ↗ |
| CodeContests | Deprecated | 44.2% | 10 | 2024-08-29 | independent | Source ↗ |
| SimpleQA | Active | 39.8% | 90 | 2024-08-29 | independent | Source ↗ |
| ComplexBench | Nearing Saturation | 57.7% | 80 | 2024-08-28 | independent | Source ↗ |
| SWE-bench Pro | Active | 42.2% | 90 | 2024-08-28 | independent | Source ↗ |
| LongBench-Pro | Active | 62.7% | 80 | 2024-08-28 | independent | Source ↗ |
| MultiPL-E | Deprecated | 62.6% | 10 | 2024-08-27 | independent | Source ↗ |
| ZeroSCROLLS | Nearing Saturation | 61.2% | 78 | 2024-08-27 | independent | Source ↗ |
| AgentIF | Nearing Saturation | 47.9% | 80 | 2024-08-26 | independent | Source ↗ |
| MATH | Saturated | 70.6% | 25 | 2024-08-26 | independent | Source ↗ |
| LiveCodeBench | Saturated | 55.8% | 39 | 2024-08-26 | independent | Source ↗ |
| InfiniteBench | Nearing Saturation | 65% | 65 | 2024-08-26 | independent | Source ↗ |
| MMLU-Pro | Nearing Saturation | 70.4% | 75 | 2024-08-25 | independent | Source ↗ |
| Mercury | Deprecated | 59.8% | 10 | 2024-08-25 | independent | Source ↗ |
| SCROLLS | Nearing Saturation | 61.2% | 65 | 2024-08-25 | independent | Source ↗ |
| RULER | Nearing Saturation | 70.4% | 84 | 2024-08-24 | independent | Source ↗ |
| TabMWP | Saturated | 66.2% | 25 | 2024-08-24 | independent | Source ↗ |
| CyberSecEval | Deprecated | 61.2% | 10 | 2024-08-24 | independent | Source ↗ |
| MMLU-Redux | Saturated | 75% | 25 | 2024-08-23 | independent | Source ↗ |
| QuALITY | Nearing Saturation | 67.3% | 78 | 2024-08-23 | independent | Source ↗ |
| SWE-bench Verified | Saturated | 53% | 25 | 2024-08-23 | independent | Source ↗ |
| MGSM | Nearing Saturation | 68.4% | 65 | 2024-08-22 | independent | Source ↗ |
| LongBench | Nearing Saturation | 65.8% | 65 | 2024-08-22 | independent | Source ↗ |
| SWE-bench | Deprecated | 46.2% | 10 | 2024-08-22 | independent | Source ↗ |
| AGIEval | Saturated | 64.5% | 25 | 2024-08-21 | independent | Source ↗ |
| QASPER | Nearing Saturation | 63.5% | 65 | 2024-08-21 | independent | Source ↗ |
| TruthfulQA | Saturated | 67.3% | 25 | 2024-08-21 | independent | Source ↗ |
| SuperGLUE | Saturated | 76.5% | 25 | 2024-08-20 | independent | Source ↗ |
| GSM-Hard | Deprecated | 69.1% | 10 | 2024-08-20 | independent | Source ↗ |
| MT-Bench | Nearing Saturation | 7.2% | 75 | 2024-08-19 | independent | Source ↗ |
| BIG-Bench Hard | Saturated | 73.5% | 25 | 2024-08-19 | independent | Source ↗ |
| CodeXGLUE | Deprecated | 63.9% | 23 | 2024-08-19 | independent | Source ↗ |
| MBPP+ | Nearing Saturation | 69.4% | 65 | 2024-08-18 | independent | Source ↗ |
| WinoGrande | Saturated | 78% | 28 | 2024-08-18 | independent | Source ↗ |
| IFEval | Deprecated | 69% | 25 | 2024-08-18 | independent | Source ↗ |
| GPQA | Nearing Saturation | 67.3% | 78 | 2024-08-18 | independent | Source ↗ |
| HumanEval+ | Saturated | 73.4% | 25 | 2024-08-17 | independent | Source ↗ |
| Needle-in-a-Haystack | Saturated | 95.3% | 44 | 2024-08-17 | independent | Source ↗ |
| CommonsenseQA | Saturated | 78.8% | 28 | 2024-08-17 | independent | Source ↗ |
| GSM8K | Saturated | 77.8% | 25 | 2024-08-16 | independent | Source ↗ |
| GLUE | Saturated | 81% | 28 | 2024-08-16 | independent | Source ↗ |
| MBPP | Saturated | 72.1% | 25 | 2024-08-16 | independent | Source ↗ |
| HumanEval | Saturated | 77.5% | 25 | 2024-08-15 | independent | Source ↗ |
| MMLU | Saturated | 78% | 25 | 2024-08-15 | independent | Source ↗ |
| ARC-AGI | Saturated | 70.5% | 25 | 2024-08-15 | independent | Source ↗ |
Local Hardware Execution (Ollama)Hardware Compatible
# 1. Pull and execute model locally with Ollama
ollama run sarvam-2b
# 2. Or invoke via local OpenAI-compatible endpoint
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "sarvam-2b",
"messages": [{"role": "user", "content": "Analyze reasoning chains on AIME 2026"}]
}'Evaluation & Sourcing Notes
Scores listed for Sarvam 2B (Sarvam-1) represent verified evaluations extracted from official research papers, independent evaluation suites (HELM, LMSYS, OpenCompass, LiveBench), and verified audit reports.
All benchmarks marked as saturated or deprecated reflect historical performance where the benchmark no longer provides active discriminative power.