Sarvam AIproprietary
Released: 2025-05-20Sarvam Vision (Document AI)
Specialized OCR and multimodal visual reasoning system designed to digitize, parse, and structure messy handwritten and printed government records, bills, and legal documents across Indian scripts.
Intelligence Index
83.5/ 100
Calibrated multi-domain compositeThroughput Speed
88tok/sec
Output streaming throughputTime To First Token
320ms
Initial chunk server latencyLiveBench Contam-Free
73.2%
September 2026 suite scoreSourced Benchmark Evaluations (61)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| ARC-AGI-3 | Active | 52.3% | 93 | 2025-06-24 | independent | Source ↗ |
| ARC-AGI-2 | Nearing Saturation | 62.8% | 75 | 2025-06-17 | independent | Source ↗ |
| FrontierMath | Active | 21% | 98 | 2025-06-14 | independent | Source ↗ |
| PutnamBench | Active | 39% | 90 | 2025-06-13 | independent | Source ↗ |
| SWE-Lancer | Deprecated | 49% | 20 | 2025-06-11 | independent | Source ↗ |
| Omni-MATH | Deprecated | 55.5% | 10 | 2025-06-11 | independent | Source ↗ |
| Soohak | Active | 57% | 98 | 2025-06-09 | independent | Source ↗ |
| FrontierCode | Active | 41.8% | 98 | 2025-06-09 | independent | Source ↗ |
| Humanity's Last Exam | Active | 35.4% | 93 | 2025-06-07 | independent | Source ↗ |
| StructFlowBench | Nearing Saturation | 65.3% | 75 | 2025-06-07 | independent | Source ↗ |
| AIME | Nearing Saturation | 71.3% | 75 | 2025-06-07 | independent | Source ↗ |
| OlympiadBench | Active | 61.5% | 80 | 2025-06-06 | independent | Source ↗ |
| MathArena | Active | 60% | 99 | 2025-06-05 | independent | Source ↗ |
| FireBench | Active | 64.5% | 95 | 2025-06-05 | independent | Source ↗ |
| Terminal-Bench | Active | 50.4% | 80 | 2025-06-05 | independent | Source ↗ |
| OpenAI-MRCR | Active | 70.1% | 80 | 2025-06-04 | independent | Source ↗ |
| SimpleQA | Active | 43.4% | 90 | 2025-06-04 | independent | Source ↗ |
| CodeContests | Deprecated | 46.8% | 10 | 2025-06-04 | independent | Source ↗ |
| ComplexBench | Nearing Saturation | 63% | 80 | 2025-06-03 | independent | Source ↗ |
| SWE-bench Pro | Active | 44.6% | 90 | 2025-06-03 | independent | Source ↗ |
| LongBench-Pro | Active | 68.5% | 80 | 2025-06-03 | independent | Source ↗ |
| ZeroSCROLLS | Nearing Saturation | 66.8% | 78 | 2025-06-02 | independent | Source ↗ |
| MultiPL-E | Deprecated | 66.2% | 10 | 2025-06-02 | independent | Source ↗ |
| LiveCodeBench | Saturated | 59% | 39 | 2025-06-01 | independent | Source ↗ |
| AgentIF | Nearing Saturation | 52.2% | 80 | 2025-06-01 | independent | Source ↗ |
| MATH | Saturated | 73.5% | 25 | 2025-06-01 | independent | Source ↗ |
| InfiniteBench | Nearing Saturation | 71% | 65 | 2025-06-01 | independent | Source ↗ |
| Mercury | Deprecated | 63.4% | 10 | 2025-05-31 | independent | Source ↗ |
| SCROLLS | Nearing Saturation | 66.8% | 65 | 2025-05-31 | independent | Source ↗ |
| MMLU-Pro | Nearing Saturation | 76.8% | 75 | 2025-05-31 | independent | Source ↗ |
| RULER | Nearing Saturation | 76.8% | 84 | 2025-05-30 | independent | Source ↗ |
| TabMWP | Saturated | 69% | 25 | 2025-05-30 | independent | Source ↗ |
| CyberSecEval | Deprecated | 64.8% | 10 | 2025-05-30 | independent | Source ↗ |
| SWE-bench Verified | Saturated | 56.2% | 25 | 2025-05-29 | independent | Source ↗ |
| MMLU-Redux | Saturated | 81.8% | 25 | 2025-05-29 | independent | Source ↗ |
| QuALITY | Nearing Saturation | 73.5% | 78 | 2025-05-29 | independent | Source ↗ |
| LongBench | Nearing Saturation | 71.8% | 65 | 2025-05-28 | independent | Source ↗ |
| SWE-bench | Deprecated | 49% | 10 | 2025-05-28 | independent | Source ↗ |
| MGSM | Nearing Saturation | 71.3% | 65 | 2025-05-28 | independent | Source ↗ |
| AGIEval | Saturated | 69.2% | 25 | 2025-05-27 | independent | Source ↗ |
| QASPER | Nearing Saturation | 69.3% | 65 | 2025-05-27 | independent | Source ↗ |
| TruthfulQA | Saturated | 73.5% | 25 | 2025-05-27 | independent | Source ↗ |
| GSM-Hard | Deprecated | 72% | 10 | 2025-05-26 | independent | Source ↗ |
| SuperGLUE | Saturated | 82.1% | 25 | 2025-05-26 | independent | Source ↗ |
| CodeXGLUE | Deprecated | 67.7% | 23 | 2025-05-25 | independent | Source ↗ |
| MT-Bench | Nearing Saturation | 7.9% | 75 | 2025-05-25 | independent | Source ↗ |
| BIG-Bench Hard | Saturated | 78.9% | 25 | 2025-05-25 | independent | Source ↗ |
| WinoGrande | Saturated | 83.7% | 28 | 2025-05-24 | independent | Source ↗ |
| MBPP+ | Nearing Saturation | 73.4% | 65 | 2025-05-24 | independent | Source ↗ |
| IFEval | Deprecated | 75.3% | 25 | 2025-05-24 | independent | Source ↗ |
| GPQA | Nearing Saturation | 73.5% | 78 | 2025-05-24 | independent | Source ↗ |
| Needle-in-a-Haystack | Saturated | 96.7% | 44 | 2025-05-23 | independent | Source ↗ |
| HumanEval+ | Saturated | 77.8% | 25 | 2025-05-23 | independent | Source ↗ |
| CommonsenseQA | Saturated | 84.5% | 28 | 2025-05-23 | independent | Source ↗ |
| MBPP | Saturated | 76.3% | 25 | 2025-05-22 | independent | Source ↗ |
| GSM8K | Saturated | 81% | 25 | 2025-05-22 | independent | Source ↗ |
| GLUE | Saturated | 86.9% | 28 | 2025-05-22 | independent | Source ↗ |
| MMLU | Saturated | 85.2% | 25 | 2025-05-21 | independent | Source ↗ |
| HumanEval | Saturated | 82.1% | 25 | 2025-05-21 | independent | Source ↗ |
| ARC-AGI | Saturated | 75.7% | 25 | 2025-05-21 | independent | Source ↗ |
| LiveBench (Continuous) | Active | 73.2% | 94 | 2025-02-01 | independent | Source ↗ |
OpenAI Compatible API ExecutionAPI Ready
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="sarvam-vision",
messages=[{"role": "user", "content": "Evaluate multi-step logic problem"}],
temperature=0.2
)
print(response.choices[0].message.content)Evaluation & Sourcing Notes
Scores listed for Sarvam Vision (Document AI) 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.