Safetensors
qwen3
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'>

[![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: []
---
```