memgui-rl/MemGUI-8B-RL

VISIONPricing:Input $0.727 / Output $5.405Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 18, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

MemGUI-8B-RL is an 8 billion parameter vision-language model based on Qwen3-VL-8B-Instruct, developed by the anonymous MemGUI-RL project. It is post-trained using FARPO (Folding-Aware Reward-Decoupled Policy Optimization) to enhance its performance as a mobile GUI agent. This model excels at complex, long-horizon mobile GUI tasks, demonstrating improved success rates on benchmarks like MemGUI-Bench and MobileWorld compared to its supervised fine-tuned predecessor.

Loading preview...

Overview

MemGUI-8B-RL is an 8 billion parameter vision-language model, building upon the Qwen3-VL-8B-Instruct architecture. Developed by the anonymous MemGUI-RL project, this model is specifically designed for long-horizon mobile GUI agent tasks. It undergoes post-training with FARPO (Folding-Aware Reward-Decoupled Policy Optimization), a technique that significantly improves its ability to handle complex interactions within mobile graphical user interfaces.

Key Capabilities

  • Enhanced Mobile GUI Automation: Achieves superior performance in automating tasks on mobile GUIs, as evidenced by its improved Pass@1, Pass@3, and IRR scores on the MemGUI-Bench benchmark (30.5 / 48.4 / 39.0 vs. 23.4 / 35.9 / 30.2 for MemGUI-8B-SFT).
  • Out-of-Distribution Task Handling: Demonstrates better generalization to unseen mobile GUI tasks, with a higher success rate on the MobileWorld GUI-only benchmark (19.7% vs. 17.9% for MemGUI-8B-SFT).
  • ConAct Interface: Operates using the ConAct (Context-as-Action) interface, allowing it to issue folding directives for its history, perform memory operations, and execute GUI actions.

Good For

  • Developing Advanced Mobile GUI Agents: Ideal for researchers and developers working on autonomous agents that interact with mobile applications over extended periods.
  • Benchmarking GUI Automation: Provides a strong baseline and improved performance for evaluating mobile GUI task completion.
  • Research in Reward-Decoupled Policy Optimization: Serves as a practical application and demonstration of the FARPO technique in a complex, real-world domain.

Finetunes
1 models