The AI Post
Agents & CodingOpen ModelsEnterpriseFundraisingGenerative MediaGovernanceInferenceInfrastructureLegal & SafetySector Impact
← Front Page Products · Alibaba · Honor

Alibaba launches Qwen Intelligence, says its 3 phone agents run on Honor

The company put end-to-end task completion at 90% and said its planner agent tops MobilePA-Bench, a benchmark suite Alibaba wrote and published itself.

Alibaba launched Qwen Intelligence on Wednesday, a package of three agents it is selling to phone makers. The three are a planner agent, a mobile-use agent that drives apps, and a creative agent that makes images. Trade reports of the launch say the stack is already running on Honor's MagicOS.

The planner agent takes first place on MobilePA-Bench, according to Alibaba, which published that benchmark alongside the launch. A technical report for Qwen-Planner-Agent 27B puts the model at 77.05% overall on it, 9.83 points above its baseline, at an estimated $2.41 per 1,000 tasks.

Alibaba put the mobile-use agent at 82.1 on MobileWorld, 92.2 on MobileWorld-Real and 97.2 on AndroidDaily, with a 90% end-to-end success rate. It said the agent calls an app's interface where one exists and falls back to reading the screen and tapping where none does.

Every figure here is Alibaba's, measured on tests it chose, and the benchmark its planner leads is one it wrote. No outside evaluator has published a score for any of the three agents. Alibaba did not give pricing for phone makers, or name other handset brands.

The creative agent produces a first image in about three seconds, which Alibaba said is roughly twice as fast as leading peers. It did not name the peers or the hardware. Honor has published no figures of its own, and the companies did not say how many phones carry the stack.

Sources 3 sources

  1. Primary Qwen
  2. Primary Qwen-Planner-Agent technical report
  3. Press KuCoin, citing BlockBeats