README.md

---
license: mit
datasets:
- cheapresearch/CheapResearch-DS-33k
---



# FlashResearch-4B-Thinking

<img src='cheap.png' width='700'>

[![Model](https://img.shields.io/badge/HF-Model-blue)](https://huggingface.co/your-username/your-model-name)
[![License](https://img.shields.io/badge/License-Apache--2.0-green)](#license)
[![Dataset](https://img.shields.io/badge/Dataset-CheapResearch--DS--33k-orange)](https://huggingface.co/datasets/cheapresearch/CheapResearch-DS-33k)

**A 4B-parameter Qwen model distilled from Tongyi DeepResearch-30B A3B**, optimized for web-scale “deep research” tasks and inference with **[Alibaba-NLP/DeepResearch](https://github.com/Alibaba-NLP/DeepResearch)**.

* **Base**: Qwen 4B (dense)
* **Teacher**: Tongyi DeepResearch 30B A3B (MoE)
* **Method**: SFT distillation on **33k** curated deep-research examples
* **Dataset**: [`flashresearch/FlashResearch-DS-33k`](https://huggingface.co/datasets/cheapresearch/CheapResearch-DS-33k)
* **Primary Use**: Fast, low-cost **DeepResearch** agent runs (browsing, multi-step reasoning, source-grounded answers)

## Evaluation 

<img src='hle.png' width='500'>
<img src='simpleqa.png' width='500'>

## Training Data

* **Primary dataset**: [`flashresearch/FlashResearch-DS-33k`](https://huggingface.co/datasets/flashresearch/FlashResearch-DS-33k)

## Inference with Alibaba-NLP/DeepResearch (Recommended)

This model is intended to be used **directly** with the DeepResearch repo.

### 1) Install & set up

```bash
git clone https://github.com/Alibaba-NLP/DeepResearch
cd DeepResearch
# Create env (example)
python -m venv .venv && source .venv/bin/activate
pip install -e .  # or pip install -r requirements.txt if provided
```

### 2) Point DeepResearch to this model

Edit the config to add this model

```bash
MODEL_PATH=flashresearch/FlashResearch-4B-Thinking
```

### Hardware notes

* **Single 12–16GB GPU** is enough for 4B FP16; FP8/INT4 quantization allows smaller VRAM. If you quantize, the summary model can be local as well.



## Acknowledgements

* Qwen team for the base 4B architecture
* Alibaba-NLP for **DeepResearch**
* CheapResearch contributors for the 33k dataset

---

## Citation

If you use this model, please cite:

```bibtex
@software{cheapresearch_thinking_2025,
  title        = {CheapResearch 4B Thinking},
  author       = {Artem Y.},
  year         = {2025},
  url          = {https://huggingface.co/flashresearch/FlashResearch-4B-Thinking}
}
```

And the dataset:

```bibtex
@dataset{cheapresearch_ds_33k,
  title        = {CheapResearch-DS-33k},
  author       = {Artem Y.},
  year         = {2025},
  url          = {https://huggingface.co/datasets/flashresearch/FlashResearch-DS-33k}
}
```

---

## Changelog

* **v1.0.0 (2025-10-04)** — First public release (33k distillation, DeepResearch-ready)



### Model Card Metadata (Hugging Face)

```yaml
---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen
- deep-research
- browsing
- citation
- reasoning
- distillation
- agent
- vllm
- cheapresearch
datasets:
- flashresearch/FlashResearch-DS-33k
base_model:
- Qwen/Qwen3-4B-Thinking-2507
model-index:
- name: FlashResearch-4B-Thinking
  results: []
---
```