Eagle2-9B

Eagle2-9B is the flagship of NVIDIA Research's Eagle 2 vision-language model family, built around a data-centric post-training philosophy rather than architectural novelty. It pairs a Qwen2.5-7B-Instruct language backbone with SigLIP and ConvNeXt vision encoders and a 16K context, and the team published a full technical report documenting the data strategies behind it.

On release the model set state-of-the-art results across document, chart, OCR, and general multimodal benchmarks, reportedly matching models with up to 70B parameters on several of them. Note the weights carry a CC-BY-NC-4.0 license, so commercial use requires separate permission; the code itself is Apache 2.0.