Qwen2 joins the multimodal race!

2-VL is a new multimodal LLM and comes in two sizes: 2B for on-device usage and 7B under Apache 2.0!

Qwen2 7B VL shows matching performance to GPT-4o mini across different benchmarks!


🧮 Comes in 2 sizes, 2B (2.2B) and 7B (8.3B) using a Vision Encoder
🎥 Can understand videos over 20 minutes for video-based question-answering
🖼️ Qwen2 7B VL around GPT-4o mini performance on VLM Benchmarks
🌍 Multilingual, including most European languages, Japanese, Korean, Arabic, and Vietnamese
📝 Improved OCR and handwritten text extraction
🤗 Available on
@huggingface

🔓 Released under Apache 2.0
🔄 Dynamic image resolutions and M-ROPE (Multimodal Rotary Position Embedding)

Blog: https://qwenlm.github.io/blog/qwen2-vl/
Models: https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d

@opendatascience