README.md
16.5 KB · 356 lines · markdown Raw
1 ---
2 library_name: transformers
3 license: apache-2.0
4 license_link: https://huggingface.co/Qwen/Qwen3-4B/blob/main/LICENSE
5 pipeline_tag: text-generation
6 base_model:
7 - Qwen/Qwen3-4B-Base
8 ---
9
10 # Qwen3-4B
11 <a href="https://chat.qwen.ai/" target="_blank" style="margin: 2px;">
12 <img alt="Chat" src="https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5" style="display: inline-block; vertical-align: middle;"/>
13 </a>
14
15 ## Qwen3 Highlights
16
17 Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features:
18
19 - **Uniquely support of seamless switching between thinking mode** (for complex logical reasoning, math, and coding) and **non-thinking mode** (for efficient, general-purpose dialogue) **within single model**, ensuring optimal performance across various scenarios.
20 - **Significantly enhancement in its reasoning capabilities**, surpassing previous QwQ (in thinking mode) and Qwen2.5 instruct models (in non-thinking mode) on mathematics, code generation, and commonsense logical reasoning.
21 - **Superior human preference alignment**, excelling in creative writing, role-playing, multi-turn dialogues, and instruction following, to deliver a more natural, engaging, and immersive conversational experience.
22 - **Expertise in agent capabilities**, enabling precise integration with external tools in both thinking and unthinking modes and achieving leading performance among open-source models in complex agent-based tasks.
23 - **Support of 100+ languages and dialects** with strong capabilities for **multilingual instruction following** and **translation**.
24
25 ## Model Overview
26
27 **Qwen3-4B** has the following features:
28 - Type: Causal Language Models
29 - Training Stage: Pretraining & Post-training
30 - Number of Parameters: 4.0B
31 - Number of Paramaters (Non-Embedding): 3.6B
32 - Number of Layers: 36
33 - Number of Attention Heads (GQA): 32 for Q and 8 for KV
34 - Context Length: 32,768 natively and [131,072 tokens with YaRN](#processing-long-texts).
35
36 For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our [blog](https://qwenlm.github.io/blog/qwen3/), [GitHub](https://github.com/QwenLM/Qwen3), and [Documentation](https://qwen.readthedocs.io/en/latest/).
37
38 > [!TIP]
39 > If you encounter significant endless repetitions, please refer to the [Best Practices](#best-practices) section for optimal sampling parameters, and set the ``presence_penalty`` to 1.5.
40
41 ## Quickstart
42
43 The code of Qwen3 has been in the latest Hugging Face `transformers` and we advise you to use the latest version of `transformers`.
44
45 With `transformers<4.51.0`, you will encounter the following error:
46 ```
47 KeyError: 'qwen3'
48 ```
49
50 The following contains a code snippet illustrating how to use the model generate content based on given inputs.
51 ```python
52 from transformers import AutoModelForCausalLM, AutoTokenizer
53
54 model_name = "Qwen/Qwen3-4B"
55
56 # load the tokenizer and the model
57 tokenizer = AutoTokenizer.from_pretrained(model_name)
58 model = AutoModelForCausalLM.from_pretrained(
59 model_name,
60 torch_dtype="auto",
61 device_map="auto"
62 )
63
64 # prepare the model input
65 prompt = "Give me a short introduction to large language model."
66 messages = [
67 {"role": "user", "content": prompt}
68 ]
69 text = tokenizer.apply_chat_template(
70 messages,
71 tokenize=False,
72 add_generation_prompt=True,
73 enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
74 )
75 model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
76
77 # conduct text completion
78 generated_ids = model.generate(
79 **model_inputs,
80 max_new_tokens=32768
81 )
82 output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
83
84 # parsing thinking content
85 try:
86 # rindex finding 151668 (</think>)
87 index = len(output_ids) - output_ids[::-1].index(151668)
88 except ValueError:
89 index = 0
90
91 thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
92 content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
93
94 print("thinking content:", thinking_content)
95 print("content:", content)
96 ```
97
98 For deployment, you can use `sglang>=0.4.6.post1` or `vllm>=0.8.5` or to create an OpenAI-compatible API endpoint:
99 - SGLang:
100 ```shell
101 python -m sglang.launch_server --model-path Qwen/Qwen3-4B --reasoning-parser qwen3
102 ```
103 - vLLM:
104 ```shell
105 vllm serve Qwen/Qwen3-4B --enable-reasoning --reasoning-parser deepseek_r1
106 ```
107
108 For local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3.
109
110 ## Switching Between Thinking and Non-Thinking Mode
111
112 > [!TIP]
113 > The `enable_thinking` switch is also available in APIs created by SGLang and vLLM.
114 > Please refer to our documentation for [SGLang](https://qwen.readthedocs.io/en/latest/deployment/sglang.html#thinking-non-thinking-modes) and [vLLM](https://qwen.readthedocs.io/en/latest/deployment/vllm.html#thinking-non-thinking-modes) users.
115
116 ### `enable_thinking=True`
117
118 By default, Qwen3 has thinking capabilities enabled, similar to QwQ-32B. This means the model will use its reasoning abilities to enhance the quality of generated responses. For example, when explicitly setting `enable_thinking=True` or leaving it as the default value in `tokenizer.apply_chat_template`, the model will engage its thinking mode.
119
120 ```python
121 text = tokenizer.apply_chat_template(
122 messages,
123 tokenize=False,
124 add_generation_prompt=True,
125 enable_thinking=True # True is the default value for enable_thinking
126 )
127 ```
128
129 In this mode, the model will generate think content wrapped in a `<think>...</think>` block, followed by the final response.
130
131 > [!NOTE]
132 > For thinking mode, use `Temperature=0.6`, `TopP=0.95`, `TopK=20`, and `MinP=0` (the default setting in `generation_config.json`). **DO NOT use greedy decoding**, as it can lead to performance degradation and endless repetitions. For more detailed guidance, please refer to the [Best Practices](#best-practices) section.
133
134
135 ### `enable_thinking=False`
136
137 We provide a hard switch to strictly disable the model's thinking behavior, aligning its functionality with the previous Qwen2.5-Instruct models. This mode is particularly useful in scenarios where disabling thinking is essential for enhancing efficiency.
138
139 ```python
140 text = tokenizer.apply_chat_template(
141 messages,
142 tokenize=False,
143 add_generation_prompt=True,
144 enable_thinking=False # Setting enable_thinking=False disables thinking mode
145 )
146 ```
147
148 In this mode, the model will not generate any think content and will not include a `<think>...</think>` block.
149
150 > [!NOTE]
151 > For non-thinking mode, we suggest using `Temperature=0.7`, `TopP=0.8`, `TopK=20`, and `MinP=0`. For more detailed guidance, please refer to the [Best Practices](#best-practices) section.
152
153 ### Advanced Usage: Switching Between Thinking and Non-Thinking Modes via User Input
154
155 We provide a soft switch mechanism that allows users to dynamically control the model's behavior when `enable_thinking=True`. Specifically, you can add `/think` and `/no_think` to user prompts or system messages to switch the model's thinking mode from turn to turn. The model will follow the most recent instruction in multi-turn conversations.
156
157 Here is an example of a multi-turn conversation:
158
159 ```python
160 from transformers import AutoModelForCausalLM, AutoTokenizer
161
162 class QwenChatbot:
163 def __init__(self, model_name="Qwen/Qwen3-4B"):
164 self.tokenizer = AutoTokenizer.from_pretrained(model_name)
165 self.model = AutoModelForCausalLM.from_pretrained(model_name)
166 self.history = []
167
168 def generate_response(self, user_input):
169 messages = self.history + [{"role": "user", "content": user_input}]
170
171 text = self.tokenizer.apply_chat_template(
172 messages,
173 tokenize=False,
174 add_generation_prompt=True
175 )
176
177 inputs = self.tokenizer(text, return_tensors="pt")
178 response_ids = self.model.generate(**inputs, max_new_tokens=32768)[0][len(inputs.input_ids[0]):].tolist()
179 response = self.tokenizer.decode(response_ids, skip_special_tokens=True)
180
181 # Update history
182 self.history.append({"role": "user", "content": user_input})
183 self.history.append({"role": "assistant", "content": response})
184
185 return response
186
187 # Example Usage
188 if __name__ == "__main__":
189 chatbot = QwenChatbot()
190
191 # First input (without /think or /no_think tags, thinking mode is enabled by default)
192 user_input_1 = "How many r's in strawberries?"
193 print(f"User: {user_input_1}")
194 response_1 = chatbot.generate_response(user_input_1)
195 print(f"Bot: {response_1}")
196 print("----------------------")
197
198 # Second input with /no_think
199 user_input_2 = "Then, how many r's in blueberries? /no_think"
200 print(f"User: {user_input_2}")
201 response_2 = chatbot.generate_response(user_input_2)
202 print(f"Bot: {response_2}")
203 print("----------------------")
204
205 # Third input with /think
206 user_input_3 = "Really? /think"
207 print(f"User: {user_input_3}")
208 response_3 = chatbot.generate_response(user_input_3)
209 print(f"Bot: {response_3}")
210 ```
211
212 > [!NOTE]
213 > For API compatibility, when `enable_thinking=True`, regardless of whether the user uses `/think` or `/no_think`, the model will always output a block wrapped in `<think>...</think>`. However, the content inside this block may be empty if thinking is disabled.
214 > When `enable_thinking=False`, the soft switches are not valid. Regardless of any `/think` or `/no_think` tags input by the user, the model will not generate think content and will not include a `<think>...</think>` block.
215
216 ## Agentic Use
217
218 Qwen3 excels in tool calling capabilities. We recommend using [Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) to make the best use of agentic ability of Qwen3. Qwen-Agent encapsulates tool-calling templates and tool-calling parsers internally, greatly reducing coding complexity.
219
220 To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.
221 ```python
222 from qwen_agent.agents import Assistant
223
224 # Define LLM
225 llm_cfg = {
226 'model': 'Qwen3-4B',
227
228 # Use the endpoint provided by Alibaba Model Studio:
229 # 'model_type': 'qwen_dashscope',
230 # 'api_key': os.getenv('DASHSCOPE_API_KEY'),
231
232 # Use a custom endpoint compatible with OpenAI API:
233 'model_server': 'http://localhost:8000/v1', # api_base
234 'api_key': 'EMPTY',
235
236 # Other parameters:
237 # 'generate_cfg': {
238 # # Add: When the response content is `<think>this is the thought</think>this is the answer;
239 # # Do not add: When the response has been separated by reasoning_content and content.
240 # 'thought_in_content': True,
241 # },
242 }
243
244 # Define Tools
245 tools = [
246 {'mcpServers': { # You can specify the MCP configuration file
247 'time': {
248 'command': 'uvx',
249 'args': ['mcp-server-time', '--local-timezone=Asia/Shanghai']
250 },
251 "fetch": {
252 "command": "uvx",
253 "args": ["mcp-server-fetch"]
254 }
255 }
256 },
257 'code_interpreter', # Built-in tools
258 ]
259
260 # Define Agent
261 bot = Assistant(llm=llm_cfg, function_list=tools)
262
263 # Streaming generation
264 messages = [{'role': 'user', 'content': 'https://qwenlm.github.io/blog/ Introduce the latest developments of Qwen'}]
265 for responses in bot.run(messages=messages):
266 pass
267 print(responses)
268 ```
269
270 ## Processing Long Texts
271
272 Qwen3 natively supports context lengths of up to 32,768 tokens. For conversations where the total length (including both input and output) significantly exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively. We have validated the model's performance on context lengths of up to 131,072 tokens using the [YaRN](https://arxiv.org/abs/2309.00071) method.
273
274 YaRN is currently supported by several inference frameworks, e.g., `transformers` and `llama.cpp` for local use, `vllm` and `sglang` for deployment. In general, there are two approaches to enabling YaRN for supported frameworks:
275
276 - Modifying the model files:
277 In the `config.json` file, add the `rope_scaling` fields:
278 ```json
279 {
280 ...,
281 "rope_scaling": {
282 "rope_type": "yarn",
283 "factor": 4.0,
284 "original_max_position_embeddings": 32768
285 }
286 }
287 ```
288 For `llama.cpp`, you need to regenerate the GGUF file after the modification.
289
290 - Passing command line arguments:
291
292 For `vllm`, you can use
293 ```shell
294 vllm serve ... --rope-scaling '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' --max-model-len 131072
295 ```
296
297 For `sglang`, you can use
298 ```shell
299 python -m sglang.launch_server ... --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}}'
300 ```
301
302 For `llama-server` from `llama.cpp`, you can use
303 ```shell
304 llama-server ... --rope-scaling yarn --rope-scale 4 --yarn-orig-ctx 32768
305 ```
306
307 > [!IMPORTANT]
308 > If you encounter the following warning
309 > ```
310 > Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'original_max_position_embeddings'}
311 > ```
312 > please upgrade `transformers>=4.51.0`.
313
314 > [!NOTE]
315 > All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, **potentially impacting performance on shorter texts.**
316 > We advise adding the `rope_scaling` configuration only when processing long contexts is required.
317 > It is also recommended to modify the `factor` as needed. For example, if the typical context length for your application is 65,536 tokens, it would be better to set `factor` as 2.0.
318
319 > [!NOTE]
320 > The default `max_position_embeddings` in `config.json` is set to 40,960. This allocation includes reserving 32,768 tokens for outputs and 8,192 tokens for typical prompts, which is sufficient for most scenarios involving short text processing. If the average context length does not exceed 32,768 tokens, we do not recommend enabling YaRN in this scenario, as it may potentially degrade model performance.
321
322 > [!TIP]
323 > The endpoint provided by Alibaba Model Studio supports dynamic YaRN by default and no extra configuration is needed.
324
325 ## Best Practices
326
327 To achieve optimal performance, we recommend the following settings:
328
329 1. **Sampling Parameters**:
330 - For thinking mode (`enable_thinking=True`), use `Temperature=0.6`, `TopP=0.95`, `TopK=20`, and `MinP=0`. **DO NOT use greedy decoding**, as it can lead to performance degradation and endless repetitions.
331 - For non-thinking mode (`enable_thinking=False`), we suggest using `Temperature=0.7`, `TopP=0.8`, `TopK=20`, and `MinP=0`.
332 - For supported frameworks, you can adjust the `presence_penalty` parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.
333
334 2. **Adequate Output Length**: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 38,912 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.
335
336 3. **Standardize Output Format**: We recommend using prompts to standardize model outputs when benchmarking.
337 - **Math Problems**: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
338 - **Multiple-Choice Questions**: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the `answer` field with only the choice letter, e.g., `"answer": "C"`."
339
340 4. **No Thinking Content in History**: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.
341
342 ### Citation
343
344 If you find our work helpful, feel free to give us a cite.
345
346 ```
347 @misc{qwen3technicalreport,
348 title={Qwen3 Technical Report},
349 author={Qwen Team},
350 year={2025},
351 eprint={2505.09388},
352 archivePrefix={arXiv},
353 primaryClass={cs.CL},
354 url={https://arxiv.org/abs/2505.09388},
355 }
356 ```