Atlas

Models

← All models

NVLM-D-72B

NVIDIA · 2024-09-30 · 79.4B parameters

NVLM-D-72B is NVIDIA's open-weight, decoder-only multimodal model in the NVLM 1.0 family; NVIDIA announced the family on September 17, 2024, and published this exact checkpoint on September 30, 2024. The complete hosted checkpoint contains exactly 79,379,593,344 parameters and combines a Qwen2-72B-Instruct language backbone, an InternViT-6B-448px-V1-5 vision encoder, and a two-layer MLP projector. It accepts text and image inputs and generates text, targeting vision-language and text-only tasks including OCR, document and chart understanding, mathematical reasoning, and coding.

Benchmark scores

BenchmarkScore
AI2D (OpenVLM)80.1
Capability118.7
CharXiv Reasoning36.8
CharXiv-D63.9
HallusionBench (OpenVLM)49.7
MathVista (OpenVLM)64.0
MM-Vet (OpenVLM)58.9
MMMU Validation (OpenVLM)60.8
MMStar (OpenVLM)63.7
ZeroBench (Main questions, pass@1)0.0
ZeroBench (Main questions, pass@5)1.0
ZeroBench (Subquestions, pass@1)14.9
ZeroBench Main Questions pass^50.0
Loading Atlas data…