File size: 3,230 Bytes
f483428 c4e1fe0 f483428 fea7dfb a1c113e f483428 f9024fe f483428 c4e1fe0 f483428 d0f0f98 f483428 d0f0f98 f483428 c4e1fe0 f483428 c4e1fe0 f483428 f9024fe f483428 5f5e3dd f483428 c4e1fe0 f483428 5f5e3dd f483428 c4e1fe0 f483428 eafdbfd f483428 c4e1fe0 f483428 119bccc f483428 c4e1fe0 f483428 d0f0f98 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 |
---
license: mit
datasets:
- cheapresearch/CheapResearch-DS-33k
---
# FlashResearch-4B-Thinking
<img src='cheap.png' width='700'>
[](https://huggingface.co/your-username/your-model-name)
[](#license)
[](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: []
---
```
|