PhoneBuddy-4B-RealApp

PhoneBuddy-4B-RealApp is the PhoneBuddy real-app-only RL ablation checkpoint without mockapps, presented in the paper PhoneBuddy: Training Open Models for Agentic Phone Use.

Project page: https://phonebuddyai.github.io/

GitHub: https://github.com/PhoneBuddyAI/phonebuddy

Model Details

• Model family: Qwen3.5 VL style checkpoint
• model_type: qwen3_5
• Processor: Qwen3VLProcessor
• Checkpoint role: real-app-only ablation, without mockapps
• Tool-call format: Qwen-style XML as defined in chat_template.jinja

The model card and repository are initially published as private for validation.

Tool-Call Format

PhoneBuddy-4B-RealApp follows the Qwen-style XML tool-call format defined by the bundled chat_template.jinja, for example:

<tool_call>
<function=example_function_name>
<parameter=example_parameter_1>
value_1
</parameter>
</function>
</tool_call>

Use the tokenizer or processor chat template from this repository when constructing prompts with tools.

Loading Environment

These checkpoints use Qwen3.5 VL style model metadata:

• model_type: qwen3_5
• Architecture: Qwen3_5ForConditionalGeneration
• Processor: Qwen3VLProcessor
• Tokenizer metadata: TokenizersBackend

Use the matching Qwen3.5 VL / PhoneBuddy training or inference environment that registers these classes. In a generic public Transformers environment, compatibility depends on whether that build includes qwen3_5 and the tokenizer backend used by this checkpoint.

A minimal processor load can be tested with:

from transformers import AutoProcessor

repo_id = "PhoneBuddyAI/PhoneBuddy-4B-RealApp"
processor = AutoProcessor.from_pretrained(repo_id, trust_remote_code=False)

Full config, tokenizer, and model loading should be done in an environment that supports the classes above. For example, public transformers==4.57.6 does not register model_type=qwen3_5, and AutoTokenizer does not import TokenizersBackend; in that environment those failures indicate version/class compatibility, not missing checkpoint files.

Intended Use

PhoneBuddy is designed for research on phone agents, multimodal tool use, and visual action reasoning. This checkpoint is intended for ablation comparisons against the main Real+Mock RL checkpoint.

Citation

@misc{tang2026phonebuddytrainingopenmodels,
 title={PhoneBuddy: Training Open Models for Agentic Phone Use},
 author={Zhengyang Tang and Xin Lai and Pengyuan Lyu and Xinyuan Wang and Tianyi Bai and Chenxin Li and Yiduo Guo and Huawen Shen and Yuxuan Liu and Junyi Li and Zhengyao Fang and Yang Ding and Yi Zhang and Weinong Wang and Xingran Zhou and Liang Wu and Fei Tang and Sunqi Fan and Shangpin Peng and Zheng Ruan and Anran Zhang and Benyou Wang and Ji-Rong Wen and Rui Yan and Chengquan Zhang and Han Hu},
 year={2026},
 eprint={2606.23049},
 archivePrefix={arXiv},
 primaryClass={cs.CL},
 url={https://arxiv.org/abs/2606.23049},
}