Alibaba's Mobile-Agent-v3.5 takes over your desktop, phone, and browser
Alibaba just dropped Mobile-Agent-v3.5 alongside the GUI-Owl-1.5 model family, spanning from 2B to 235B parameters. Built on Qwen3-VL, it natively unifies vision, reasoning, and tool execution across desktop, mobile, and browser environments, topping 20+ benchmarks. Model weights are now open on Hugging Face with live demos available.

The biggest headache with AI automation has never been understanding the prompt. It is the friction of jumping between screens. Data found on a browser has to be manually pasted into desktop software, while mobile-only apps remain trapped on your phone. Alibaba wants to erase those boundaries entirely.
Alibaba has released and open-sourced Mobile-Agent-v3.5, a cross-platform GUI agent powered by the new GUI-Owl-1.5 foundation model family. Built on top of Qwen3-VL, the lineup spans from a nimble 2B parameters all the way to 235B, offering both Instruct and Thinking variants. Across more than 20 benchmark evaluations, it achieved state-of-the-art results.
Bridging phone, PC, and browser
Most existing automation tools are siloed. They are either browser extensions, desktop script wrappers, or mobile accessibility hacks. Mobile-Agent-v3.5 takes an end-to-end multimodal route, natively unifying perception, element grounding, planning, and execution across mobile, desktop, and web environments.
In practical demos, the agent chains cross-application tasks seamlessly. You can ask it to look up Apple and Nvidia stock prices in Microsoft Edge, and it will autonomously open WPS Office, create a spreadsheet, and log company names in column A with prices in column B.
In another test, it checked airfare from Guangzhou to Chengdu on Ctrip, grabbed the cheapest option, checked train fares on the same route, and compared prices. It even hopped across Xiaohongshu and Douyin to tally up follower counts across platforms.
Under the hood, the system features long-horizon memory and supports Model Context Protocol (MCP) tool invocation, letting the model figure out when to click a pixel and when to call an API.
Getting your hands on it
The release is already widely accessible. Model weights are hosted on Hugging Face alongside technical reports and codebases. If you do not want to set up local environments, live demos and inference APIs are available on ModelScope and Alibaba Cloud Bailian. It is also integrated into Alibaba's Wuying Cloud Phone, giving developers a ready-to-use virtual Android environment without needing physical hardware.
Why it matters
We think this release signals a major shift from specialized point tools to unified operating system agents. What Alibaba did right here is the architectural spread. Having 2B and 4B models means fast, lightweight on-device execution is within reach, while the 235B heavyweight can handle complex reasoning in the cloud.
For builders, this drastically lowers the bar for automating real-world workflows that span across phones and PCs without brittle platform-specific glue code. The hard problem with computer-use agents has always been error recovery—one misclick on a dynamic UI can derail an entire task. But by consolidating cross-platform vision, reasoning, and tool use into a single stack, Alibaba is pushing GUI agents much closer to daily utility.
Sources (1)
- github.com(opens in a new tab)Originalgithub.com