For the past two years, GUI agents — the kind that look at a screen and click buttons — have been held back by two things: interactive environments are expensive, and realistic training tasks are scarce. Frontier models have already harvested the gains of online reinforcement learning with verifiable rewards in math and code, but that playbook breaks down on GUI tasks. You can't auto-grade a math proof the way you can grade an arithmetic answer, and it's hard to run hundreds of thousands of trajectories across dozens of real phones at once.
MAI-Gym, released by Alibaba's Tongyi MAI team, tries to fuse evaluation, data collection, and training into a single pipeline so both bottlenecks get addressed at once.
The key design: one rollout pipeline shared by eval and training
This is the platform's most consequential trade-off. If the environment you evaluate in and the environment you train in are two separate codebases, then a score you measure today and a model you train tomorrow are always separated by a layer of drift that can never be reconciled. MAI-Gym collapses both into one rollout pipeline, so the agent behaves consistently across evaluation and training.
On top of that, the platform supports ten established benchmarks and scales to 20,000 parallel environment instances. For anyone doing agent training, that number matters more than it first appears — the strength of online RL is a direct function of how many interactive environments you can run at once.
Not just infrastructure: synthesizing apps and tasks
Parallel environments alone aren't enough, because annotating realistic tasks is brutally expensive. So the platform does something else — it automatically synthesizes GUI applications and hard tasks, including resettable mocked apps and coherent cross-app worlds where a task stays consistent across multiple applications, backed by execution-validated task verifiers. Whether a task succeeded is decided by actually running it, not by a human-written label.
The paper notes that task generation is guided by the model's weaknesses. The release includes 49 simulated apps, 2,000 execution-verified tasks, and reusable synthesis skills.
MAI-Gym-9B: one model for phone and desktop
Trained on this platform with joint online RL, the team produced a unified 9B model, MAI-Gym-9B, that handles both mobile and desktop operation. The paper reports 78.4% on MobileWorld and 76.2% on OSWorld-Verified, and claims state-of-the-art results among open GUI agents under 10B parameters. That claim comes from the authors' own evaluation; no independent team has reproduced it yet.
A more interesting number comes from the transfer experiment: RL trained inside a mocked train-ticketing app improved real-phone task success on its live counterpart from 28.8% to 48.1%. That suggests training signal synthesized in a mock environment does cross the gap between simulation and reality.
My take: the GUI agent bottleneck is shifting from models to environment engineering
The past two years of industry narrative has almost entirely focused on the model side — parameters, context length, multimodal capability. But the real blocker for GUI agents is quietly moving to the other side of the equation: how you build environments, how you verify tasks, how high you can scale parallelism.
MAI-Gym packages those engineering problems into one open platform and releases the weights alongside it, so nobody has to rebuild those details from scratch to stand up their own agent pipeline. Simulation-to-real transfer remains the biggest open question — 28.8% to 48.1% shows transfer genuinely works, but it also shows how limited things were beforehand: more than half the tasks failed outright on a real phone.
So the interesting part of this release isn't "another 9B model posting numbers." It's that the open-source community now has a complete loop joining evaluation, synthesis, training, and verification. Paper, code, and weights are all out (details on the Hugging Face paper page). What's worth watching is how many people can build their own agents on this foundation.
When verifiable rewards spread from math problems to every click and every swipe, agent training has only just begun to scale.