Qwen2-VL: dynamic resolution, 20+ minute video and open weights for 2B and 7B
In one sentence On August 29, 2024 Alibaba's Qwen team unveils Qwen2-VL: images at arbitrary resolution, understanding of videos over 20 minutes, agent capabilities. 2B and 7B ship with Apache 2.0 open weights, 72B via API only.
Qwen2-VL is the new vision-language model family from Alibaba's Qwen team. The model handles images of any resolution by turning them into a variable number of visual tokens, instead of resizing everything to a fixed size. It understands videos over 20 minutes long and reads text in images in many languages beyond English and Chinese.
It can also act as an agent: starting from what it sees and from text instructions, it can operate a phone or control a robotic arm, and it supports function calling.
At launch the 2 and 7 billion parameter versions are downloadable under the Apache 2.0 license; the largest version, with 72 billion parameters, is available only through an API.
Companies
Alibaba, Qwen Team
Tools
Qwen2-VL, Qwen2-VL-2B, Qwen2-VL-7B, Qwen2-VL-72B
Tags
Sources