Alibaba Tongyi-MAI Releases OpenGCUA and MAI-Gym: Full Open Stack for Generalist Computer Use Boosts OSWorld-v2 by +19.6 Points and Cuts Steps by 53%

Background and the Problem

Computer-use agents are transitioning from brittle GUI-only screen clicking to Generalist Computer-Use Agents (GCUAs) that seamlessly interleave visual GUI actions with programmatic code execution. However, frontier systems such as GPT-6 Astra's computer-use harness remain closed-source, and scaling online reinforcement learning (RL) for GUI agents has been severely bottlenecked by slow environment instantiation and scarce verifiable tasks.

Architecture and How It Works

Alibaba's Tongyi-MAI team open-sourced two complementary stacks—OpenGCUA (arXiv:2610.22200) and MAI-Gym (arXiv:2610.20939):

1. OpenGCUA (Code + GUI Unification & Harness-to-Model Transfer): Empowers agents to dynamically combine programmatic execution with visual GUI actions while shifting observation and task-state decisions from rigid harness rules to the model, backed by a real-device harness and cross-device SDK;
2. MAI-Gym (20,000 Parallel RL Environments & Mocked App Synthesis): Unifies evaluation and online RL across mobile, desktop, and web environments at up to 20,000 parallel instances, featuring automated synthesis of resettable mocked apps with deterministic execution verifiers.

Benchmarks and Measured Results

Extensive evaluations across OSWorld-v2, MobileWorld, and physical macOS/Android hardware demonstrate massive gains:

• +19.6 Points and 53.0% Fewer Steps on OSWorld-v2: OpenGCUA improves scores across all eight tested models (including GPT-6 Astra and Claude Opus 5), raising mean partial score by +19.6 points and reducing average steps per task by 53.0%, while revealing that Qwen3.8-Max consumes 19.4x more reasoning tokens than GPT-6 Astra on real macOS workflows;
• MAI-Gym-9B Hits 78.4% MobileWorld & 76.2% OSWorld-Verified: Sets a new open-weight SOTA under 10B parameters, while demonstrating Sim-to-Real transfer on a physical smartphone where mocked train-ticketing RL boosted real-app success from 28.8% to 48.1%.

Getting Started for Developers

Both repositories, trajectories, and checkpoints are available at https://github.com/Tongyi-MAI/OpenGCUA and https://github.com/Tongyi-MAI/MAI-Gym:

• Adopt Hybrid GUI + Code Harnesses: Integrate OpenGCUA's cross-device SDK so desktop agents can bypass multi-step UI clicking whenever a single shell or Python command suffices;
• Bootstrap GUI RL in Mocked Worlds: Train domain-specific mobile/web agents on MAI-Gym's resettable mocked apps before transferring policies to production apps.