开源手机智能体OpenMobile-2发布 性能追平闭源模型

Summary 1: OpenMobile-2 is an open mobile-agent project jointly released by Nanjing University, SenseTime, and other partners; its core appeal is that the model, data, and training environment are all open.

1. It offers 9B- and 27B-parameter models based on Qwen3.5 and Qwen3.6, with performance benchmarked against proprietary closed-source models.

2. It reconstructs 35 mainstream apps and covers more than 110 apps in its training environment, releasing the full training environment, datasets, and reproduction instructions.

Summary 2: Practical and performance details center on benchmark results and experimental features.

1. In AndroidWorld, the 27B version rises from 67.1 to 79.9 and the 9B version from 54.3 to 74; in MobileWorld, the 27B version rises from 35.9 to 50.4; in SPA-Bench, the 27B version rises from 31.9 to 59.6.

2. More than 50 simulated apps expose direct command interfaces; after training, task completion rates improve by up to 10.3 percentage points and average five fewer steps. However, these interfaces exist only in simulated apps and remain experimental.

Summary 1: Although the article does not directly address brand marketing or pricing, brands can focus on product R&D, user behavior, and cross-app experiences.

1. OpenMobile-2 reconstructs 35 apps including Taobao, JD.com, Didi, Meituan, Slack, and Google Drive, and its training environment contains more than 110 apps, indicating that high-frequency services and overseas tools are important scenarios for mobile agents.

2. The agent sees only screenshots and cannot read internal app data, closely mirroring human operation logic; the environment can retain orders, budgets, and travel plans across sessions and allows clipboard transfer across apps.

Summary 2: The implications for brand and product teams lie in interaction entry points and user journeys.

1. More than 50 simulated apps expose direct command interfaces, allowing actions such as adding items to cart to bypass layered menus; after training, task completion rates improve by up to 10.3 percentage points, with an average of five fewer steps per task.

2. App diversity has a greater impact on performance than total data volume: increasing the number of covered apps raises success rates on unseen-app tasks from 60.7% to 67.5%, suggesting brands should focus on multi-scenario adaptability and entry-point design.

Summary 1: Sellers can see automation and partnership opportunities in consumer scenarios through the open-sourcing of mobile agents.

1. OpenMobile-2 opens its models, training environment, datasets, and reproduction instructions; the 9B and 27B models enable low-cost experimentation. Training covers high-frequency services such as Taobao, JD.com, Didi, and Meituan, and orders, budgets, and travel plans can persist across sessions.

2. More than 50 simulated apps expose direct command interfaces, so actions such as adding items to cart can be triggered directly; after training, task completion rates improve by up to 10.3 percentage points, with an average of five fewer steps.

Summary 2: Opportunities come with risk caveats.

1. Direct command interfaces currently exist only in the team's simulated apps, and the functionality remains experimental; it should not be treated as a deployed platform capability.

2. Reinforcement learning rewards actual task completion and penalizes false claims of completion. Base training plus reinforcement learning yields only limited gains in similar scenarios: in AndroidWorld, the 27B version slips slightly from 80.8 to 79.9 after reinforcement learning. Sellers should pay attention to real-task validation and scenario-transfer risk.

Summary 1: Factories can draw digitalization and product-design lessons from the open models and training environment.

1. The project releases 9B and 27B models, a full training environment, datasets, and reproduction instructions, lowering the barrier to experimentation. The training environment reconstructs 35 mainstream apps; together with simulated programs and Android-emulator apps, it covers more than 110 apps.

2. The agent obtains only screenshots and cannot read internal app data. The environment can retain orders, budgets, and travel plans, supports clipboard transfer across apps, and resets automatically after completion, reflecting a reproducible, resettable simulation approach.

Summary 2: Commercial opportunities lie in scenario adaptation and interface design.

1. At the same data scale, app diversity has a greater impact on performance than total data volume; increasing the number of covered apps raises success rates on unseen-app tasks from 60.7% to 67.5%.

2. More than 50 simulated apps expose direct command interfaces; after training, task completion rates improve by up to 10.3 percentage points and average five fewer steps, with transfer to untried tools. This offers a reference for product design and digital process reengineering, though the interfaces remain experimental.

Summary 1: Service providers should focus on mobile-agent training pain points and open-source solutions.

1. Real app logins, real transactions, and non-resettable states make training costly and workflows hard to reproduce. Frontier training data is largely controlled by large tech companies, and open-source models have long lagged closed-source ones.

2. OpenMobile-2 reconstructs 35 mainstream apps; including simulated programs and Android-emulator apps, it covers more than 110 apps and releases the environment, datasets, and reproduction instructions. Using Gemini 3.1 Pro as the teacher model, it generates about 11,600 example workflows and more than 2,000 practice tasks with automatic success validation.

Summary 2: Technical trends and customer value can be illustrated by benchmarks and experimental features.

1. AndroidWorld, MobileWorld, and SPA-Bench all improve; the 27B version rises from 31.9 to 59.6 on SPA-Bench. App diversity is more critical than total data volume.

2. Direct command interfaces can reduce click steps, raising task completion rates by up to 10.3 percentage points. However, reinforcement learning gains are limited: the 27B version on AndroidWorld slips slightly from 80.8 to 79.9 after reinforcement learning. Service providers need to prioritize scenario matching and real validation.

Summary 1: Marketplace sellers should pay attention to new requirements for platform capabilities and operational rules from mobile agents.

1. Agent training is constrained by real logins, real transactions, and non-resettable states; agents can see only screenshots and cannot read internal app data. OpenMobile-2 uses 35 reconstructed apps and an environment of more than 110 apps, supports cross-session retention of orders, budgets, and travel plans, enables clipboard transfer across apps, and automatically resets after simulation.

2. More than 50 simulated apps expose direct command interfaces that bypass layered menus; after training, task completion rates improve by up to 10.3 percentage points and average five fewer steps. However, the interfaces exist only in simulated apps and remain experimental.

Summary 2: Operations management and risk avoidance should emphasize validation and transfer.

1. Task results are determined by programmatic comparison of app storage data before and after execution, not by a language model. It builds a 215-task benchmark, nearly half of which require cross-app completion, and verifies low similarity between training and test tasks.

2. Reinforcement learning rewards actual completion and penalizes false claims. Base training plus reinforcement learning yields only limited gains in similar scenarios; the 27B version's AndroidWorld score slips slightly from 80.8 to 79.9. Platforms need to guard against false completion and scenario-generalization risk.

Summary 1: This research offers new directions for open-sourcing mobile agents and training paradigms.

1. OpenMobile-2 is based on Qwen3.5 9B and Qwen3.6 27B, releases models, training environment, datasets, and full reproduction instructions, and benchmarks against proprietary closed-source models. It reconstructs 35 apps, covers more than 110 apps in its environment, and uses Gemini 3.1 Pro to generate about 11,600 workflows and more than 2,000 automatically validated tasks.

2. Benchmarks show AndroidWorld rising from 67.1 to 79.9 for the 27B version and from 54.3 to 74 for the 9B version; MobileWorld rising from 35.9 to 50.4 for the 27B version; and SPA-Bench rising from 31.9 to 59.6 for the 27B version.

Summary 2: The study also raises new questions for discussion.

1. At the same data scale, app diversity has a greater impact than total data volume, raising success rates on unseen-app tasks from 60.7% to 67.5%. Base training plus reinforcement learning yields limited gains; the 27B version on AndroidWorld slips slightly from 80.8 to 79.9.

2. Direct command interfaces raise task completion rates by up to 10.3 percentage points, reduce steps by an average of five, and can transfer. The study builds a 215-task benchmark, nearly half of which are cross-app, uses programmatic data comparison to determine results, and verifies low similarity. Related work includes Microsoft's API vs. GUI classification, OPPO X-OmniClaw, and Google Gemini 3.5 Flash embedded in phone control.