Alibaba Releases Qwen-UI-Agent with 12 Mobile Actions Without Bash
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
The Qwen-UI-Agent repository provided by Alibaba exposes twelve actions focused on mobile interface and does not offer Bash execution. Despite presenting a unified GUI+CLI action space with grouped actions, the available code operates on a single action per turn and does not support grouping.
The code only supports one action per turn and no Bash execution
In the available version, the system only handles one action at a time, and a regular expression is used to prevent any additional calls to a tool. Mentions of bash or shell in the repository are limited to benchmark images or documentation, with no examples of executable actions. Bash or CLI execution is not available in the downloadable code. Even using the "escape hatch" MCP mechanism, it is not possible to group multiple actions, and action grouping is not supported in the provided environment.
What is actually delivered: a mobile tool and 12 gestures
The repository provides twelve actions, all limited to screen gestures or management functions. The only tool offered, "mobile_use," groups these twelve actions that do not involve the shell. No shell interface is included. An MCP mechanism allows for the attachment of external tools, including shell-type tools, without modifying the repository, but this does not change the limitation to one action per turn or the lack of action grouping.
GUI+CLI promise and positioning against benchmarks
Alibaba presents Qwen-UI-Agent as a hybrid agent unifying GUI and CLI, with grouped actions, and claims support for Bash actions. The agent was recently showcased, and its performance is reported on seven benchmarks. However, it appears that the agent is primarily a renamed wrapper of MAI-UI, a mobile agent that only allows one action per turn, and some features announced in the documentation are not present in the provided checkpoints. It is recommended to directly consult the prompt and action definitions in the repository to know the exact number of supported actions before any integration.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.