SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Mano - A GUI intelligent operation model launched by Mininglamp Technology

Author: AI小集 · Original source: AI工具集

What is Mano

Mano is a proprietary large model launched by Mininglamp Technology, focused on intelligent operation of graphical user interfaces (GUI). The model is based on a multimodal foundation model and, through innovations such as online reinforcement learning and automatic training data collection, has achieved SOTA results on the two major benchmarks Mind2Web and OSWorld. Mano can accurately identify and operate GUI elements in web pages and desktop environments, completing complex tasks such as filling out forms and logging into accounts, providing an efficient solution for automated operations and advancing the development of the GUI agent field.

Mano

Mano's main features

  • Automated web operations: The model can automatically complete various operations on web pages, such as filling out forms, clicking buttons, entering text, and submitting forms, and can be applied in scenarios such as automated data collection and web automation testing.
  • Desktop application operations: Supports operations on desktop software, including opening software, performing menu operations, entering text, and clicking buttons.
  • Cross-platform operations: Compatible with multiple operating systems and browsers, enabling automated operations on different platforms and meeting diverse automation needs.
  • Data collection and analysis: Supports automatically collecting data from web pages or desktop applications and performing preliminary analysis, providing support for subsequent data processing and decision-making.
  • Error detection and recovery: Equipped with an error detection mechanism, it can promptly detect errors during operations and attempt automatic recovery, improving the reliability and stability of operations.

Mano's technical principles

  • Multimodal foundation model: Based on a multimodal foundation model, it can understand and process visual information (such as web page screenshots) and textual information (such as user instructions and web page text), enabling perception and understanding of the GUI environment.
  • Online reinforcement learning: Through online reinforcement learning, Mano can continuously learn and optimize operational strategies through interaction with the real environment, improving adaptability and decision-making ability in dynamic environments.
  • Automatic training data collection: A training data automatic collection module is designed to automatically generate and collect high-quality interaction data for model training and optimization, reducing manual annotation costs.
  • Supervised fine-tuning (SFT): In the first stage of training, supervised learning is used to fine-tune the model, enabling it to better understand and execute specific GUI operation tasks.
  • Offline reinforcement learning: In the second stage, offline reinforcement learning further optimizes the model's decision-making ability, enabling it to better complete multi-step operation tasks.
  • Online reinforcement learning: In the third stage, the model interacts with the real environment in a simulated environment, and through online reinforcement learning further improves the model's adaptability and flexibility.
  • Verification module: Mano is equipped with a verification module used to verify the correctness of each operation step, promptly detect and correct errors, and improve the accuracy and reliability of operations.

Mano's project address

  • Technical paper：https://www.mininglamp.com/wp-content/uploads/2025/09/%E6%98%8E%E7%95%A5%E7%A7%91%E6%8A%80-Mano-Technical-Report.pdf

Mano's application scenarios

  • Automated data collection: Mano can automatically collect data from web pages or desktop applications, providing efficient support for data analysis and research and saving manual collection time.
  • Web automation testing: The model automatically completes various web operations, such as filling out forms, clicking buttons, etc., used to test whether web functions are working properly, improving testing efficiency and accuracy.
  • Enterprise office automation: Supports operations on desktop software to automate internal enterprise office processes, such as automatically filling out reports, sending emails, etc., improving work efficiency.
  • Software automation testing: The model is used for automated testing of desktop software, including opening software, performing menu operations, etc., helping developers quickly discover software issues.
  • Intelligent customer service assistance: The model can automatically handle some common customer inquiry issues, such as checking order status, answering common questions, etc., reducing the workload of customer service staff.