infinitylogesh/GUI-Decisions-31B

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

GUI-Decisions-31B is a 31 billion parameter vision-language model based on google/gemma-4-31B-it, specifically fine-tuned for computer-use decision-making. It utilizes a unique slot-reading LoRA to interpret actions, coordinates, and other UI elements directly from a single forward pass, significantly accelerating decision grounding compared to traditional decoding methods. This model excels at tasks requiring precise interaction with graphical user interfaces, such as identifying click locations, scroll actions, and object bounding boxes, making it ideal for automated UI control and agentic applications.

Loading preview...

Overview

GUI-Decisions-31B is a 31 billion parameter vision-language model built upon google/gemma-4-31B-it, enhanced with a specialized slot-reading LoRA. This model is designed to make computer-use decisions by reading actions, coordinates (256 bins per axis), swipe endpoints, keys, and scroll parameters directly from next-token distributions at fixed slots, rather than through sequential decoding. This method allows for significantly faster grounding of UI interactions.

Key Capabilities

  • Accelerated UI Decision-Making: Achieves up to 3.5x faster computer-use grounding by reading decision parameters from a single forward pass, reducing latency from 472-546 ms to 146-210 ms per step.
  • Precise UI Interaction: Identifies specific actions (e.g., click, scroll, press), precise screen coordinates, and bounding boxes for general objects (center + log size) and coarse masks (center + 24 rays).
  • Efficient Deployment: Can be served with vLLM for high-throughput inference, with a median step time of 190 ms on an RTX PRO 6000 GPU.
  • Object Localization: Capable of identifying and boxing objects in images, achieving a mean IoU of 0.770 on RefCOCO val photo boxes and 0.744 on COCO val2017 boxes.

Good For

  • Automated UI Control: Ideal for agents that need to interact with graphical user interfaces quickly and precisely.
  • Robotic Process Automation (RPA): Automating tasks that involve navigating and manipulating software applications.
  • Computer Vision for UI: Tasks requiring the identification of interactive elements and their properties on a screen.
  • Research in Efficient LLM Grounding: Exploring novel methods for faster and more direct interpretation of visual information for decision-making.

Finetunes
1 models