A3-Qwen3.5-9B
| 💾 Code | 📄 Paper | 🌐 Website |
|---|---|---|
| 🤗 Dataset | 🤖 Models | 📦 PyPI |
Structured Distillation of Web Agent Capabilities Enables Generalization
Xing Han Lù, Siva Reddy
A3-Qwen3.5-9B is a 9B multimodal web agent fine-tuned from Qwen/Qwen3.5-9B on A3-Synth, a synthetic dataset generated using the Agent-as-Annotators (A3) framework.
The model achieves 41.5% on WebArena, surpassing closed-source models such as Claude 3.5 Sonnet (36.0%) and GPT-4o (31.5%) under the same evaluation protocol.
Usage
Serve with vLLM:
vllm serve McGill-NLP/A3-Qwen3.5-9B --tensor-parallel-size 2 --max-model-len 65536 --enforce-eager --dtype bfloat16
Training
- Base model: Qwen/Qwen3.5-9B
- Data: A3-Synth (16k examples from Gemini 3 Pro trajectories)
- Method: SFT with FSDP
- Max sequence length: 16,384
- Learning rate: 1e-5
- Epochs: 2
- Batch size: 1 per GPU, gradient accumulation 4
Model Variants
| Model | Parameters | Link |
|---|---|---|
| A3-Qwen3.5-9B | 9B | McGill-NLP/A3-Qwen3.5-9B |
| A3-Qwen3.5-4B | 4B | McGill-NLP/A3-Qwen3.5-4B |
| A3-Qwen3.5-2B | 2B | McGill-NLP/A3-Qwen3.5-2B |
Citation
@misc{lu2026structured,
title={Structured Distillation of Web Agent Capabilities Enables Generalization},
author={Xing Han Lù and Siva Reddy},
year={2026},
eprint={2604.07776},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2604.07776},
}