Alibaba's Qwen team has released Qwen2.5-VL, a new family of AI models designed to work across text, images, documents, videos and software interfaces. The launch arrives while DeepSeek is drawing heavy attention in the tech industry, but Alibaba is using Qwen2.5-VL to show that its own AI work is moving quickly in multimodal systems.
The models are not only built to answer questions about content. According to the source, they can parse files, understand videos, count objects in images and interact with software on PCs and mobile devices. That last capability places Qwen2.5-VL in the same broad category as systems that try to operate apps and computers on a user's behalf.
What Qwen2.5-VL is built to do
Qwen2.5-VL is presented as a vision-language model family with a wide set of analysis tasks. The Qwen team says the models can analyze charts and graphics, extract data from scans of invoices and forms, and "comprehend" multiple-hours-long videos.
Those abilities matter because many practical AI tasks are not limited to clean text prompts. Business documents, forms, charts, screenshots and long videos all contain information that a useful assistant may need to interpret. In that sense, Qwen2.5-VL is aimed at the messy mix of formats people actually work with.
The source also says Qwen2.5-VL can recognize "IPs from film and TV series, as well as a wide variety of products," per the team. The article notes that this suggests the models might have been trained in part on copyrighted works.
For users and developers, the immediate access points are Alibaba's Qwen Chat app and Hugging Face, where the models are available to download. That gives the release both a consumer-facing testing path and a developer-facing distribution channel.
Benchmarks put the flagship model in direct comparison
Per the Qwen team's benchmarking, the strongest Qwen2.5-VL model beats OpenAI's GPT-4o, Anthropic's Claude 3.5 Sonnet and Google's Gemini 2.0 Flash on a range of evaluations. The listed areas include video understanding, math, document analysis and question-answering.
Benchmarks are not the same as real-world reliability, but they frame how Alibaba wants the model family to be judged. The comparisons also show that Qwen2.5-VL is being positioned against some of the best-known AI systems from major companies.
The flagship model is Qwen2.5-VL-72B. The source describes it as the largest and most capable model in the family. It is also the model used in the example where Qwen Chat declined to respond to a politically sensitive prompt.
Software control is the most eye-catching feature
One of the most important claims around Qwen2.5-VL is that it can interact with software on PCs and mobile devices. The source compares this to the model powering OpenAI's recently launched Operator.
A video posted on X by Philipp Schmid, a technical lead at Hugging Face, showed Qwen2.5-VL launching the Booking.com app for Android and booking a flight from Chongqing to Beijing. That example is significant because it shows a model moving beyond content analysis into a sequence of app actions.
Another video showed a Qwen2.5-VL model controlling apps on a Linux desktop. In that case, however, the source says it did not appear to accomplish much beyond switching tabs.
The same caution appears in the benchmark picture. Qwen's benchmarking shows Qwen2.5-VL scoring poorly on OSWorld, a benchmark that tries to mimic a real computer environment. That makes the software-control story more complicated: the model can perform some visible actions, but the source does not support the idea that it is already broadly reliable at operating a computer.
Restrictions remain part of the product experience
Because Qwen2.5-VL is AI developed by a Chinese company, the source says it has certain restrictions on the topics it will discuss, at least in Qwen Chat. When the source article's author asked Qwen2.5-VL-72B to talk about "Xi Jinping's mistakes," Qwen Chat produced an error message.
The source also states that China's internet regulator benchmarks many models developed in the country to ensure their responses "embody core socialist values." Many Chinese AI systems decline to respond to topics that might raise regulatory concerns, such as Taiwan's autonomy.
That means users testing Qwen2.5-VL through Qwen Chat may encounter limits that are not simply technical. The model family can be evaluated on analysis tasks, app control and benchmark scores, but the product experience is also shaped by topic restrictions.
Licensing divides the model family
The Qwen2.5-VL series includes smaller models and a flagship model with different licensing terms. The two smaller, less sophisticated models, Qwen2.5-VL-3B and Qwen2.5-VL-7B, are available under a permissive license.
The largest model, Qwen2.5-VL-72B, is under Alibaba's custom license. That license requires companies and developers with more than 100 million monthly active users to request permission from Qwen/Alibaba before deploying the model commercially.
This split matters for adoption. Developers can experiment with smaller models under more permissive terms, while large-scale commercial deployment of the flagship model comes with an additional permission requirement.
Overall, Qwen2.5-VL shows Alibaba pushing multimodal AI toward a broader role: reading files, interpreting images, understanding video and taking actions inside software. The release is ambitious, but the source also points to limits, especially around real computer use and restricted topics in Qwen Chat.