ServiceNow/GroundNext-7B-V0

GroundNext-7B-V0: Vision-Language Model for GUI Grounding

GroundNext-7B-V0, developed by ServiceNow, is a 7 billion parameter vision-language model specifically designed for GUI element grounding within the GroundCUA project. It is built on the Qwen2.5-VL-7B-Instruct base and features a two-stage training approach (SFT + RLOO) using only 700K human-annotated desktop demonstrations, significantly less data than prior work.

Key Capabilities & Performance

• Superior Grounding Accuracy: Achieves 52.9% on ScreenSpot-Pro, 67.7% on OSWorld-G, and 60.3% on UI-Vision benchmarks, outperforming larger models like UI-TARS-72B.
• Exceptional Cross-Platform Generalization: Despite desktop-only training, it shows strong generalization with 81.1% accuracy on MMBench-GUI and 90.4% on ScreenSpot-v2.
• Agentic Capabilities: When paired with reasoning models (e.g., OpenAI o3), it reaches a 50.6% overall success rate on OSWorld for end-to-end computer use.
• Native Tool-Calling Support: Integrates built-in computer use action space for mouse, keyboard, and screen interactions.
• High Context Length: Inherits a 262,144 token context length from its base model.

Good For

• Automating Desktop GUI Interactions: Ideal for tasks requiring precise identification and interaction with graphical user interface elements.
• Developing Computer Use Agents: Its agentic capabilities and tool-calling support make it suitable for building intelligent agents that can operate software.
• Research in GUI Grounding: Provides a strong baseline and efficient training methodology for further research in vision-language models for computer control.

Limitations

• Primarily trained on desktop environments, though it generalizes well.
• Currently supports mouse click actions only.
• Optimized for English UI elements.