Rumiii/Qwen3-0.6B-BrowserAgent

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Rumiii/Qwen3-0.6B-BrowserAgent is a compact 0.8 billion parameter browser agent model, fine-tuned from Qwen/Qwen3-0.6B with LoRA. It is specifically designed to output precise browser actions like 'click('37')' based on a task goal and BrowserGym page observations. This model excels at automating web interactions, demonstrating significant improvements in exact match accuracy (from 0.83% to 89.58%) and target element matching (from 4.17% to 92.5%) compared to its base model, making it ideal for controlled browser automation tasks.

Loading preview...

Overview

Rumiii/Qwen3-0.6B-BrowserAgent is a specialized, compact browser automation model built upon the Qwen3-0.6B architecture. It is fine-tuned using LoRA on the saital/browser-agent-phase1-sft-action-only dataset, which consists of successful MiniWoB steps generated by a larger teacher model. The model's primary function is to take a task goal and a BrowserGym page observation as input and output only the next precise action, such as click('37').

Key Capabilities

  • Precise Action Generation: Outputs specific browser actions like clicks and text fills based on visual observations.
  • High Accuracy: Achieves an 89.58% exact match accuracy, a substantial improvement from the base model's 0.83%.
  • Reliable Action Types: Demonstrates 100% action type matching and 100% valid format compliance.
  • Target Element Identification: Shows a 92.5% accuracy in matching target elements for actions.
  • Efficient Operation: A compact 0.8 billion parameter model, suitable for deployment where resource efficiency is important.

Good For

  • Automated Web Interaction: Ideal for tasks within controlled environments like BrowserGym or MiniWoB.
  • Synthetic Task Execution: Excels at performing tasks similar to those it was trained on (around 30 synthetic MiniWoB task types).
  • Action-Only Output: Useful for systems requiring direct, unambiguous browser commands without additional reasoning text.

Limitations

  • Limited Real-World Transfer: Performance on real-world websites is not expected to be strong due to training on synthetic data.
  • No Error Recovery: Trained only on successful trajectories, it cannot recover from mistakes.
  • Weak on Ambiguous Elements: Struggles with tasks requiring selection among many similar elements or exact text copying.

Finetunes
1 models