Published:
Author:
Published:
Author:

Announcing Proxy Lite: A mini, open-weights, version of our Proxy assistant.

We’re excited to announce Proxy Lite – a 3B parameter VLM that brings state of the art web automation capabilities to the open source community. Our WebVoyager results place Proxy Lite among the top performers in web automation tasks, while using just a fraction of the computational resources.

Proxy Lite also comes with a comprehensive framework for VLM-browser interaction giving enterprise grade browser control. You can try it out here!

How it works

Proxy Lite responses are done in 3 distinct steps allowing for better generalization than traditional prompt-prediction models.

Each response is separated into three parts:

Using execution feedback similar to DeepSeek R1, Proxy Lite learns to observe and reason, letting it advance on a wide range of tasks.

Here you can see a more detailed breakdown of our task success rates. For even more detail check out our full trajectories on clusterfudge.

Website NameSuccess Rate (%)Finish Rate (%)Avg. Messages
Allrecipes87.895.110.3
Amazon70.095.07.1
Apple82.189.710.7
ArXiv60.579.116.0
BBC News69.477.815.9
Booking70.085.024.8
Cambridge Dictionary86.097.75.7
Coursera82.597.54.7
ESPN53.897.514.9
GitHub85.092.510.0
Google Flights38.551.334.8
Google Map78.994.79.6
Google Search71.492.96.0
Huggingface68.674.318.4
Wolfram Alpha78.393.56.1

You can find the version of WebVoyager we used here.

Next Steps

We are excited to release this work to the open source community. While Proxy Lite is already the best open weights agent, we’re just getting started.

Proxy Lite shows that a smart training regime can compete with massive parameter counts. We invite developers and researchers to join us – whether by contributing code, building applications, or sharing use cases.

Stay tuned for updates, and don’t forget to star our repository!

Share

Sign Up for the Enterprise Waitlist

Enterprise Form