Back to directory

Mobile-Agent

A family of GUI agent models and systems from Tongyi Lab for automating desktop, mobile and browser interfaces, built on GUI-Owl foundation models with tool and MCP calling and long-horizon memory.

Mobile-Agent

Mobile-Agent is a family of GUI agent models and systems from Tongyi Lab, Alibaba Group, for automating desktop, mobile and browser interfaces. The underlying GUI-Owl model family (available in 2B, 4B, 8B, 32B and 235B sizes, in instruct and thinking variants) is built on Qwen3-VL and reports state-of-the-art results across more than 20 GUI benchmarks, covering end-to-end tasks, grounding, tool and MCP calling, and long-horizon memory.

ToolCUA is a related end-to-end computer-use agent that orchestrates when to take a GUI action versus invoke a tool, trained with a two-stage pipeline of trajectory-aware tool synthesis followed by online agentic reinforcement learning. Mobile-Agent-v3.5 is available through Alibaba Cloud's Wuying Cloud Phone, a cloud-based Android environment, and OSWorld-MCP is a companion benchmark for evaluating MCP tool-invocation capability in realistic scenarios.

It can be tried through hosted demos and APIs on ModelScope and Alibaba Cloud Bailian without deploying models locally, and model weights are published on Hugging Face. It is written in Python under the MIT license.

More in Agent frameworks

See all 902
  • Recurse

    recurse.run

    A serverless harness that lets a coding agent design, test, and deploy its own specialist sub-agents as tools, MCP servers, or bots, defined by a schema-validated manifest.

  • ProofForge

    shulgin.is-a.dev

    An AI-agent pipeline that decomposes a mathematics problem, proves each piece, and formalizes the result in Lean 4 and Mathlib until the Lean kernel accepts it, with six of its proofs already merged into Google DeepMind's formal-conjectures repository.

  • local-coder

    github.com

    Builds a local coding-agent team on top of Ollama and OpenCode by detecting the machine's hardware, recommending models for seven specialist roles, and testing whether each assigned model can actually use tools, edit a repository and pass verification.