NVLM-D-72B

NVLM-D-72B is the flagship of NVIDIA's NVLM 1.0 family, released September 2024. It is a 72B-parameter multimodal model that handles charts, tables, memes, images, math, and code. The unusual finding in NVIDIA's paper is that multimodal training improved text-only performance: accuracy rose by an average of 4.3 points across text benchmarks versus the LLM backbone, reaching 81.7 on MMLU and 93.2 on GSM8K. On vision-language tests the team reports 58.7/54.9 on MMMU, 65.2 on MathVista, and 852 on OCRBench, competitive with GPT-4o on several of them; treat these as team-reported. Weights are public on Hugging Face with a promise of training code to follow. The license is CC-BY-NC 4.0, which bars commercial use, so it is a research and evaluation artifact rather than a production base for companies. It remains a useful reference for how multimodal training can strengthen rather than dilute a language model.

Team-reported: MMMU 58.7/54.9, MathVista 65.2, OCRBench 852; text: MMLU 81.7, GSM8K 93.2, MATH 73.1