Open GUI Agent Platform Scales RL Training to 20,000 Instances, Ships 9B Checkpoint

A research team from Alibaba Group has released MAI-Gym, an open platform designed to close the gap between evaluating GUI agents and training them with reinforcement learning at scale — a gap that has kept serious RL-based agent training largely inside large, well-resourced labs. The platform runs across mobile, computer, and web environments simultaneously, and includes a 9B-parameter model checkpoint that the authors say achieves state-of-the-art performance among open GUI agents with fewer than 10 billion parameters, with a 78.4% success rate on the MobileWorld benchmark and 76.2% on OSWorld-Verified.

The more consequential result in the paper may not be those benchmark scores, which are author-reported from a preprint submitted October 14, 2026, and have not yet been independently reproduced. It is the sim-to-real transfer measurement: RL training conducted entirely inside a synthesized, mocked version of a train-ticketing app pushed the agent's real-phone success rate from 28.8% to 48.1% — a 19-percentage-point gain on a physical Android device that never participated in training. That result offers a direct answer to a structural question that has shadowed GUI agent deployment: whether performance gains earned in controlled simulation carry through to the inconsistency and variation of real applications.

The Infrastructure Problem MAI-Gym Is Built to Solve

Training GUI agents with reinforcement learning against real applications is operationally difficult in ways that training language models is not. Real apps change their layouts between updates, require authenticated accounts, enforce rate limits, and can produce side effects — placing an actual order, sending an actual message — that cannot be easily undone. Running tens of thousands of RL training episodes against a live banking app or a real ticketing system is not safe at scale.

The standard workaround has been offline learning from curated demonstrations, but offline methods hit an effectiveness ceiling: the agent learns to imitate recorded behavior rather than to reason about novel interface states. Online RL, where the agent runs inside an environment and updates its weights from interaction-derived rewards, consistently outperforms offline imitation in robotics and game-playing. The question for GUI agents has been whether a sufficiently realistic training environment can be built without using live production apps.

MAI-Gym's answer is a synthesis pipeline that generates mocked GUI applications — functional interfaces with controllable, resettable internal state — alongside an automated task verifier that checks completion deterministically rather than relying on an LLM judge. The synthesized apps are not screenshots or static mockups; they are interactive environments that the agent can navigate, and whose state can be reset between episodes without any persistent side effects. The paper also describes cross-app worlds: coherent multi-app task environments where an agent must, for example, open a contact from one application and then compose a message in another, with correct task verification spanning both.

The architecture that ties the system together is a shared rollout pipeline — the same execution environment is used for both evaluation runs and RL training runs. This is not incidental engineering. When evaluation and training environments diverge, agents optimized in one often perform differently in the other; combining them removes a source of distribution shift that can mislead training signals. Most previous systems used separate infrastructure for evaluation and training, which required researchers to either accept that gap or invest in keeping two codebases synchronized.

A Platform Built for 20,000 Parallel Environments

Scale is where MAI-Gym departs most sharply from previous open efforts. The paper reports that the system supports up to 20,000 simultaneous environment instances. A directly comparable open platform, MobileGym, published in May 2026 by researchers at a separate institution, describes a single server hosting "hundreds" of parallel mobile instances — a useful capability, but one that differs from 20,000 by roughly two orders of magnitude.

The scale difference matters for RL training in particular. Reinforcement learning algorithms that work well for GUI agents — gradient-based policy optimization methods — improve sample efficiency when they can average over large batches of rollout experience. A platform that can run thousands of simultaneous agents collecting environment interactions produces usable training signal far faster than one running dozens. The practical effect is that a research team without a dedicated cluster of real devices or a proprietary simulation stack can now attempt the kind of RL sweeps that were previously limited to organizations with substantial infrastructure.

The platform supports ten established benchmarks as evaluation targets, covering mobile and computer-use scenarios, which means researchers training on MAI-Gym's synthesized environments can measure progress against standard external test sets without switching infrastructure.

Read more: New Benchmark Shows Top AI Models Score Below 20% on Agent System Building

MAI-Gym-9B Scores in Context: Author-Reported Results and the Independent Leaderboard Gap

The MAI-Gym-9B model is a unified agent for both mobile and computer environments, trained using the platform's RL infrastructure. The authors report 78.4% on MobileWorld and 76.2% on OSWorld-Verified. Both figures come from the preprint and have not yet been submitted to or independently reproduced on public leaderboards.

On OSWorld-Verified, the current highest-scoring open-weight model below 10B parameters is UI-Mate-9B, developed by Tencent, at 66.2%. If MAI-Gym-9B's 76.2% holds under independent evaluation, it would represent a roughly 10-percentage-point improvement over the current published open leader in that size class. The closest open models on OSWorld-Verified overall are UI-Mate-27B (Tencent) at 77%, and Holo3-35B-A3B (H Company) at 82.6%, both substantially larger.

On the MobileWorld leaderboard, Qwen3.8 Max, Alibaba's largest open GUI agent, stands at 77.8% as of mid-October 2026. MAI-Gym-9B's 78.4% claim would put it fractionally ahead — a striking result for a model nine times smaller in parameter count than most of the top-tier competition. MobileWorld does not yet have broad leaderboard coverage, and the benchmark organizers note sparse independent harness coverage, so cross-model comparisons should be treated with caution until the evaluation protocol is more widely standardized.

It is worth noting that the MAI-Gym team and the authors of the MAI-UI agent family share substantial overlap — eight of the eleven MAI-UI co-authors appear among the fifteen MAI-Gym authors, with senior researcher Steven C.H. Hoi leading both projects at Alibaba Group. MAI-UI, a GUI agent family ranging from 2B to 235B parameters developed by Alibaba's Tongyi Lab, previously achieved 76.7% on AndroidWorld. The two platforms represent complementary output from the same broader research organization rather than independent parallel efforts.

Benchmark Discipline: What MobileWorld and OSWorld-Verified Actually Test

GUI agent benchmarks have proliferated rapidly, and the differences between them matter for interpreting any result.

MobileWorld evaluates agents performing interactive tasks on a simulated smartphone environment. Tasks are long-horizon and require the agent to navigate real-app-style interfaces across multiple steps. The benchmark explicitly tests mobile interaction workflows rather than question-answering or web browsing, which makes it relevant for evaluating agents intended to automate smartphone use. The benchmark is relatively new and has sparse independent leaderboard coverage, meaning published scores are largely self-reported by model developers.

OSWorld-Verified is a curated subset of the OSWorld benchmark, which tests agents performing open-ended tasks on a full desktop computer environment, including file management, web browsing, and application interaction. "Verified" refers to a subset with higher-confidence automatic verification. OSWorld has broader independent coverage — benchlm.ai lists 37 models — though 34 of those entries are provider self-reports rather than independently run evaluations. The full leaderboard is dominated by closed-weight frontier models: Qwen3.8 Max (Alibaba) leads at 86.1%, followed by Claude Fable 5 and Claude Mythos 5 at 85%.

Neither benchmark has a standardized independent harness that runs every model under identical conditions, which means comparing across models requires accepting that evaluation environments and settings may not have been perfectly matched. MAI-Gym-9B's scores should be interpreted as an author's evaluation of their own model on these benchmarks until independent reproduction occurs.

Sim-to-Real Transfer: The Result That Matters for Deployment

The train-ticketing transfer experiment is the most operationally grounded result in the paper. Researchers trained the agent inside a mocked version of a train-ticketing application — a synthesized interface with controllable state and deterministic task verification — using the MAI-Gym RL pipeline. They then evaluated the trained agent on a physical Android phone running the actual train-ticketing app, without any additional fine-tuning on real-device data.

Before RL training, the agent succeeded on real-phone train-ticketing tasks 28.8% of the time. After training in the mocked environment, success on the real phone reached 48.1%. That is a 19.3-percentage-point improvement from training the agent had never seen a real device during.

MobileGym, the competing mobile-only platform, reports a comparable sim-to-real result: their GRPO-trained agent retained 95.1% of simulation-side training gains when evaluated on real devices, using a different base model and a smaller number of parallel environments. The two results, from independent teams using different methodologies, both point in the same direction: training RL GUI agents in well-constructed synthetic environments produces meaningfully better real-device performance, even when the agent never sees the real app during training.

This finding carries a practical implication for enterprise deployment. Organizations considering GUI agents for task automation — customer-service workflows, data-entry pipelines, multi-step procurement processes — have faced a training-data problem: real-environment RL is operationally risky, and demonstration data can only go so far. A validated sim-to-real methodology opens the possibility of training agents against controllable mocked versions of internal tools before deploying them against production systems.

Read more: RL Training Time for Robot Foundation Models Cut Up to 68.5% in NVIDIA-KAIST Research

Limitations and What the Numbers Do Not Yet Confirm

A 48.1% real-phone success rate on a specific task type is not production-grade reliability. For workflows where incomplete or incorrect task execution causes meaningful harm — financial transactions, account modifications, communication sent on behalf of a user — a system failing roughly half the time requires substantial human oversight or fallback infrastructure. The MAI-Gym paper establishes proof of concept for sim-to-real transfer but does not characterize how performance generalizes across different app types, task complexity levels, or interface variations.

The benchmark scores are from a preprint submitted October 14 and have not been submitted to public leaderboards or evaluated by independent parties. Scores from self-evaluated preprints have historically sometimes shifted when reproduced under controlled conditions by different teams. Until MAI-Gym-9B appears on the benchlm.ai or equivalent leaderboards under a consistent evaluation protocol, the claimed rankings within the open sub-10B category should be treated as estimates rather than established facts.

The CC BY-NC-SA 4.0 license on the released model checkpoint restricts commercial use. Organizations wanting to build commercially deployed GUI agents on the MAI-Gym-9B weights would need a separate licensing arrangement. As with any open-weight model from a Chinese research organization, researchers evaluating MAI-Gym for deployment in sensitive enterprise contexts should apply standard due-diligence practices regarding data handling and jurisdictional considerations.

The paper lists 15 co-authors from Alibaba Group, with Steven C.H. Hoi, currently a Vice President at Alibaba, serving as senior author. Full per-author institutional sub-affiliations were not available from the arXiv abstract page at the time of publication.

A Field Moving From Demonstrations to Live Interaction

GUI agent research has followed a trajectory that mirrors the broader arc of deep learning: early systems were trained mostly on curated human demonstrations, which set a ceiling on performance tied to the quality and coverage of the recorded data. The shift to online RL — letting agents learn from direct environment interaction — has been the obvious next step for years, but it has been constrained by the lack of safe, scalable training environments.

AndroidWorld, published in 2024, introduced a standardized task set for mobile agents but did not include a training infrastructure. OSWorld added computer-use evaluation benchmarks. Neither addressed the RL training environment problem directly. MAI-Gym's contribution is less about adding a new benchmark and more about building the scaffolding that lets RL training happen at the scale where it becomes competitive with fine-tuned closed-weight frontier models — or, in the sub-10B open category, potentially ahead of them.

Open Release: Framework, Tasks, and Model Checkpoint

MAI-Gym's open release includes the platform framework, the mocked-app synthesis tools, 2,000 execution-verified tasks, and the MAI-Gym-9B model checkpoint. This combination matters more than any individual component. A benchmark plus a model is a point comparison. A training platform plus a synthesis pipeline plus a verified task set is infrastructure that other researchers can extend.

The absence of a publicly available, high-scale GUI agent RL training platform has meant that most RL-based GUI agent research has either worked at small scale (tens to hundreds of environments), used proprietary internal infrastructure not available to the broader community, or avoided online RL entirely in favor of offline methods. MAI-Gym changes that constraint for researchers with access to the compute needed to run its environments, though 20,000 parallel instances is still a substantial hardware requirement.

The most significant near-term test will be whether independent teams can reproduce MAI-Gym-9B's benchmark scores using the released infrastructure, and whether the mocked-app synthesis pipeline proves general enough to extend to application domains beyond the train-ticketing use case demonstrated in the paper. If the sim-to-real results replicate across a wider range of app types, the methodology described in MAI-Gym represents a practical path toward training production-grade GUI agents without requiring direct access to live production systems during training.