PatronusAI/Qwen3.8-27B-Figmatrace-SFT

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 19, 2026License:cc-by-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

PatronusAI/Qwen3.8-27B-Figmatrace-SFT is a 27 billion parameter vision-language model developed by Patronus AI, fine-tuned from Qwen/Qwen3.8-27B. This model specializes in next-action prediction for GUI agents, trained on expert Figma design workflows to predict Playwright-MCP-style tool calls from screenshots and instructions. It excels at generalizing to various GUI tasks, outperforming its base model and some frontier models on specific benchmarks for mobile, web, and desktop navigation. Its primary strength lies in precise GUI interaction and design-related automation.

Loading preview...

Overview

PatronusAI/Qwen3.8-27B-Figmatrace-SFT is a 27 billion parameter vision-language model developed by Patronus AI, fine-tuned from Qwen/Qwen3.8-27B. It is designed for next-action prediction in GUI agents, taking a screenshot and an instruction to emit Playwright-MCP-style tool calls with norm-1000 coordinates. The model was trained on the FigmaTrace dataset, which captures expert Figma design workflows using phase-based segmentation.

Key Capabilities

  • GUI Agent Automation: Predicts the next GUI action for tasks across mobile, web, and desktop environments.
  • Design Workflow Specialization: Optimized for design-related tasks, leveraging insights from expert Figma sessions.
  • Improved Performance: Outperforms its base model on six of seven benchmark configurations, notably on AndroidControl (+16.4) and GUI-Odyssey (+9.7).
  • Precise Element Selection: Demonstrates significantly reduced error in element selection and coordinate understanding compared to the base model.
  • Norm-1000 Coordinates: Emits actions using resolution-independent norm-1000 coordinates for consistent targeting.

When to Use This Model

  • Direct Use: Ideal for building GUI agents that require precise next-action prediction based on visual input and instructions, particularly for design, web, or mobile automation.
  • Downstream Fine-tuning: Serves as a strong starting checkpoint for further supervised fine-tuning or reinforcement learning on various computer-use tasks, including design-specific agents for component construction, layout, or accessibility.

Limitations

  • Regression on Undirected Tasks: Shows reduced performance on open-ended exploration tasks like undirected VideoGUI.
  • Repetition Issues: Can sometimes emit near-identical coordinates on consecutive steps, potentially straining agent context budgets.
  • Screen-Center Bias: May prioritize content in the middle of the frame, potentially due to its FigmaTrace training.
  • Design Taste Specificity: Open-ended design tasks reflect the preferences of the experts recorded in the training data.

Finetunes
1 models