convergence-ai/proxy-lite-3b
The convergence-ai/proxy-lite-3b is a 3 billion parameter Vision-Language Model developed by Convergence AI, fine-tuned from Qwen/Qwen2.5-VL-3B-Instruct. This model functions as a web-browsing agent, designed to automate tasks within a web browser environment. It excels at web-based automation, achieving a 72.4% success rate on the WebVoyager benchmark, making it a leading open-weights model for this specific application.
Loading preview...
Proxy Lite: A Mini Web-Browsing Vision-Language Agent
Proxy Lite, developed by Convergence AI, is a 3 billion parameter Vision-Language Model (VLM) specifically engineered for automated web browsing tasks. Fine-tuned from Qwen/Qwen2.5-VL-3B-Instruct, this model acts as a web-browsing agent, capable of interacting with web environments to complete specified tasks.
Key Capabilities
- Web Automation: Designed and trained to automate routine tasks within a web browser.
- Vision-Language Integration: Processes both visual (screenshots) and textual information from web pages to make decisions.
- Tool Use: Integrates with browser tools to perform actions like navigation and data extraction.
- Context Management: Efficiently manages message history and visual observations, discarding non-current image tokens to optimize context window usage while retaining awareness of the full interaction history.
- High Performance on WebVoyager: Achieved a 72.4% success rate on the WebVoyager benchmark, positioning it as the top-performing open-weights model for this task.
Good for
- Automating repetitive web-based workflows: Ideal for tasks such as data collection, form filling, or navigating complex websites.
- Developing autonomous web agents: Provides a robust foundation for building applications that require intelligent interaction with web interfaces.
- Research in web-scale AI and agentic models: Offers an open-weights, mini version of Proxy for experimentation and development.
Proxy Lite is licensed under CC-BY-NC-4.0 and is not intended for high-stakes, safety-critical, unauthorized data extraction, or legally sensitive applications.