---
license: apache-2.0
base_model:
  - Qwen/Qwen3.5-4B
library_name: transformers
pipeline_tag: text-generation
tags:
  - medical
  - agent
  - tool-use
  - web-search
  - reinforcement-learning
---

# MedSearch-R1

MedSearch-R1 is a locally deployable medical search agent policy initialized from Qwen3.5-4B. The model is trained through cold-start knowledge distillation, step-level on-policy distillation, trajectory-level on-policy distillation, and accuracy-based reinforcement learning.

This repository contains the policy-model weights and tokenizer only. The Search--Visit tools, source-policy filters, helper-model configuration, and evaluation pipeline are not embedded in the checkpoint. Exact agent-loop code and reproducibility configurations will be provided in the associated GitHub repository.

## Loading

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "xiaohan666/MedSearch-R1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
```

The checkpoint uses the Qwen3.5 architecture and requires a Transformers release with `Qwen3_5ForCausalLM` support.

## Intended Use

The model is intended for research on medical reasoning agents and tool-augmented language models. Reproducing the paper's agent results requires the accompanying Search--Visit loop and source-filtering configuration.

## Limitations

MedSearch-R1 is not a medical device and must not be used as a substitute for professional medical judgment. Generated answers and retrieved evidence can be incomplete or incorrect. Local policy inference reduces full-context exposure to external model providers, but generated search queries may still reveal medical concepts and do not constitute a formal privacy guarantee.

## License

The model is released under the Apache License 2.0, following the license of the Qwen3.5-4B base model.