YangC777/AGW-35B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 30, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

AGW-35B is a 35.1 billion parameter screenshot-based GUI agent developed by YangC777, fine-tuned from Qwen3.5-35B-A3B. It is trained on 79,266 synthetic interaction trajectories from AutoGUIWorld, covering Chrome, Ubuntu, Windows, and macOS environments. This multimodal Mixture-of-Experts model predicts direct computer use actions from task instructions, screenshots, and interaction history, excelling in cross-platform desktop and browser task execution with significant performance gains over its base model on various GUI agent benchmarks.

Loading preview...

Overview of AGW-35B

AGW-35B is a 35.1 billion parameter multimodal GUI agent, fine-tuned by YangC777 from the Qwen3.5-35B-A3B base model. It specializes in interpreting screenshots and task instructions to predict direct computer use actions across multiple operating systems and applications. The model was trained on 79,266 synthetic GUI interaction examples generated by AutoGUIWorld, a system that expands training coverage without requiring execution in real software environments.

Key Capabilities

  • Cross-Platform GUI Automation: Proficient in automating tasks across Chrome, Ubuntu, Windows, and macOS environments, including browser, office, creative, system, and scientific workflows.
  • Screenshot-Based Action Prediction: Given a task instruction, current screenshot, and interaction history, it predicts the next executable computer_use action.
  • Direct-Action Agent: Operates as a non-thinking agent, outputting short action descriptions followed by structured tool calls with resolution-independent 0-999 coordinate grids.
  • Enhanced Performance: Demonstrates significant gains over its base model on benchmarks like OSWorld (+7.8), Windows Agent Arena (+8.5), macOSWorld (+16.9), and ScienceBoard (+18.2), with an overall mean score improvement of +12.8.
  • Improved Visual Grounding: Achieves a +25.4 gain on the ScreenSpot-Pro benchmark, particularly strong in text targets (+30.7).

Good For

  • Research and Development of GUI Agents: Ideal for exploring and building screenshot-based agents.
  • Cross-Platform Task Execution: Suitable for automating tasks in diverse desktop and browser environments.
  • Action Grounding: Effective in complex professional interfaces where precise action grounding is required.
  • Evaluating Synthetic GUI Trajectories: Useful for post-training evaluation of agents using synthetic data.
  • Agent Systems: Can be integrated into systems that execute model-produced actions within controlled environments, requiring an external executor and environment management.

Finetunes
1 models