Qwen2-VL-7B

Qwen2-VL-7B is a 7-billion-parameter multimodal large language model from the Qwen team at Alibaba. It was released on August 29, 2024 as part of the Qwen2-VL series alongside a 2B variant, with a 72B model following later.

The model upgrades image comprehension, video understanding up to 20 minutes long, and multilingual text-image processing over Qwen-VL. It handles arbitrary image resolutions through Naive Dynamic Resolution and uses Multimodal Rotary Position Embedding (M-ROPE) to process text and visual data together. Benchmarks at release put it close to GPT-4o and Claude 3.5 Sonnet on vision tasks. It ships under the Apache 2.0 license.