NVLM: benchmark results
NVIDIA's NVLM 1.0 decoder-only vision-language release, combining Qwen2-72B-Instruct with an InternViT vision encoder for multimodal and text tasks. Its card distinguishes Hugging Face and Megatron evaluation results, so those execution paths are not assumed interchangeable. Provider: NVIDIA. Released 2024-09-17. Access: Open.
Aggregate Rating 1438 ± 30, rank #4266 of 10916 rated models, from 545 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MEGA-Bench Task - Media QA Web Stack Overflow | 78.6 | Task Score (%) | 100 |
| MEGA-Bench Task - CVBench Adapted CVBench Count | 71.4 | Task Score (%) | 94.2 |
| MEGA-Bench Task - Medical Multi Organ Segmentation Rater | 42.9 | Task Score (%) | 94.2 |
| MEGA-Bench Task - FUNSD Document QA | 85.7 | Task Score (%) | 90.7 |
| MEGA-Bench Task - Table Understanding Complex Question Answering | 64.3 | Task Score (%) | 90.7 |
| MEGA-Bench Task - MMSoc Memotion | 67.1 | Task Score (%) | 89.5 |
| MEGA-Bench Task - Vibe Eval | 65.7 | Task Score (%) | 88.5 |
| MEGA-Bench Task - Red Teaming Visual Misleading | 90 | Task Score (%) | 88.4 |
| MEGA-Bench Task - Geometry Transformation | 28.6 | Task Score (%) | 87.2 |
| MEGA-Bench Task - Handwritten Math Expression Extraction | 78.6 | Task Score (%) | 87.2 |
| MEGA-Bench Task - MFC Bench Check Out Of Context | 85.7 | Task Score (%) | 86.5 |
| MEGA-Bench Task - Red Teaming Visual Order B | 90.5 | Task Score (%) | 86 |
Interactive version: theaggregate.ai/model?slug=nvlm · How It Works · Data refreshed daily, snapshot 2026-11-08.