duvoai/duvo-eye-1.5

duvo-eye-1.5: Enhanced GUI Grounding Model

duvo-eye-1.5 is a 35B-A3B (approximately 3 billion active parameters) Vision-Language Model from Duvo AI, specialized in single-step GUI element grounding. Given a screenshot and a natural-language target description, it outputs a precise {"x", "y"} click position within a [0, 1000] normalized range. This model is the grounding component of a larger computer-use stack, designed to resolve where to interact based on a planner's what.

Key Capabilities & Features

• Precise Single-Step Grounding: Outputs exact click coordinates for GUI elements, refined through GRPO reinforcement learning.
• Efficiency: Operates with ~3B active parameters, making it highly efficient for single-forward-pass inference.
• Improved Performance: Shows small but consistent gains across public grounding benchmarks (ScreenSpot-Pro, OSWorld-G, UI-I2E-Bench) compared to its predecessor, duvo-eye-1, with 0% parse failures.
• Reliable Output: Designed to directly emit JSON coordinates when enable_thinking=False is properly configured, avoiding reasoning text.

Intended Use Cases

• GUI Automation: Ideal for desktop automation pipelines, mapping textual descriptions to click points on screenshots.
• Grounding Stage: Serves as a robust grounding layer behind separate planning and verification components in agentic systems.
• Enterprise UIs: Particularly effective for professional software and enterprise back-office interfaces, where it was extensively trained and refined.

Important Considerations

• Not an Agent: duvo-eye-1.5 is a grounder, not an agent; it does not perform planning, navigation, or multi-step execution.
• No Abstention: It always returns a coordinate, even if the target is absent, requiring pipeline-level handling for absence.
• Weakness in Icons: Performance is stronger on text targets (82%) than on small icon targets (60%).
• Disable Thinking: Crucially, enable_thinking=False must be set during inference to ensure direct coordinate output.