Alibaba launches Qwen Intelligence, an AI agent kit for phone makers

Alibaba's Qwen team has announced Qwen Intelligence, which phone makers can add to their own AI assistants. It comes with three agents: one that plans tasks, one that operates the phone and one that creates images.

Paylaş
Alibaba launches Qwen Intelligence, an AI agent kit for phone makers

Alibaba's Qwen team has announced Qwen Intelligence, an AI solution for phone makers. Manufacturers can add it to their system assistants and AI services to give their phones an assistant that understands a request and carries out the task itself. The service is offered through Alibaba Cloud and billed by usage.

What Qwen Intelligence is

According to Qwen's docs, the kit targets problems phone makers face: finding the right models for phone tasks, orchestrating agents reliably, dealing with inconsistent interfaces across apps like contacts, calendar and gallery, and balancing personalisation with privacy. Instead of building these parts themselves, manufacturers plug ready-made agents into their products. Some work runs on the phone and some in the cloud, and the system routes each task accordingly.

Three agents

  • Mobile Planner Agent: Breaks a goal such as "arrange my business trip next week" into steps, plans transport, hotel and calendar, and runs other agents and apps in order. It remembers preferences the user allows, and the user can view and delete that memory.
  • Mobile-Use Agent: Reads the screen and taps, types and swipes on a real phone. Examples in the docs include checking a train time and adding a meeting to the calendar, or finding a café on a map and comparing its reviews. It pauses for confirmation before sensitive steps such as payments or sending information.
  • Creative Agent: Generates images from a one-line request, for uses like changing a photo's background in the gallery or making wallpapers and stickers. It comes in a light and a standard version.

Open benchmarks

Qwen also open-sourced the benchmarks it uses on GitHub. MobilePA-Bench measures how well phone agents plan complex tasks, and MobileWorld measures how they handle real apps and cross-app workflows. The planner agent's framework is on GitHub as Qwen-Planner-Agent.

Things to keep in mind

  • Qwen Intelligence is a paid cloud service for phone makers and developers. End users will only see these agents once a manufacturer builds them into its phones.
  • The docs are currently in Chinese, and the example scenarios use Chinese apps (12306, DingTalk, Amap, Dianping, Xiaohongshu).
  • For a fuller picture, follow evaluations from independent researchers as well. For image generation time, Qwen's docs give about 6 seconds for simple requests and about 10 seconds for complex ones.

Sources: Qwen Intelligence website, Qwen Intelligence docs, GitHub: Qwen-Planner-Agent, MobilePA-Bench, GitHub: MobileWorld.

Devamını oku

Mistral تطلق نموذج Voxtral مفتوحا لتحويل الكلام العربي إلى نص فوريا

Mistral تطلق نموذج Voxtral مفتوحا لتحويل الكلام العربي إلى نص فوريا

أطلقت Mistral AI نموذج Voxtral Mini 4B Realtime Arabic الذي يحول الكلام بالعربية ولهجاتها إلى نص في الوقت الفعلي، بترخيص Apache 2.0. وإلى جانبه نموذج LIDstral Arabic الذي يحدد لهجة النص المكتوب بالحروف العربية أو لغته. وقد طُوّر النموذجان بالتعاون مع وزارة الانتقال الرقمي المغربية.

GlobalFeed Editor tarafından