README.md
16.2 KB · 351 lines · markdown Raw
1 ---
2 library_name: transformers
3 license: apache-2.0
4 license_link: https://huggingface.co/Qwen/Qwen3-32B/blob/main/LICENSE
5 pipeline_tag: text-generation
6 ---
7
8 # Qwen3-32B
9 <a href="https://chat.qwen.ai/" target="_blank" style="margin: 2px;">
10 <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;"/>
11 </a>
12
13 ## Qwen3 Highlights
14
15 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:
16
17 - **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.
18 - **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.
19 - **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.
20 - **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.
21 - **Support of 100+ languages and dialects** with strong capabilities for **multilingual instruction following** and **translation**.
22
23 ## Model Overview
24
25 **Qwen3-32B** has the following features:
26 - Type: Causal Language Models
27 - Training Stage: Pretraining & Post-training
28 - Number of Parameters: 32.8B
29 - Number of Paramaters (Non-Embedding): 31.2B
30 - Number of Layers: 64
31 - Number of Attention Heads (GQA): 64 for Q and 8 for KV
32 - Context Length: 32,768 natively and [131,072 tokens with YaRN](#processing-long-texts).
33
34 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/).
35
36 ## Quickstart
37
38 The code of Qwen3 has been in the latest Hugging Face `transformers` and we advise you to use the latest version of `transformers`.
39
40 With `transformers<4.51.0`, you will encounter the following error:
41 ```
42 KeyError: 'qwen3'
43 ```
44
45 The following contains a code snippet illustrating how to use the model generate content based on given inputs.
46 ```python
47 from transformers import AutoModelForCausalLM, AutoTokenizer
48
49 model_name = "Qwen/Qwen3-32B"
50
51 # load the tokenizer and the model
52 tokenizer = AutoTokenizer.from_pretrained(model_name)
53 model = AutoModelForCausalLM.from_pretrained(
54 model_name,
55 torch_dtype="auto",
56 device_map="auto"
57 )
58
59 # prepare the model input
60 prompt = "Give me a short introduction to large language model."
61 messages = [
62 {"role": "user", "content": prompt}
63 ]
64 text = tokenizer.apply_chat_template(
65 messages,
66 tokenize=False,
67 add_generation_prompt=True,
68 enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
69 )
70 model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
71
72 # conduct text completion
73 generated_ids = model.generate(
74 **model_inputs,
75 max_new_tokens=32768
76 )
77 output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
78
79 # parsing thinking content
80 try:
81 # rindex finding 151668 (</think>)
82 index = len(output_ids) - output_ids[::-1].index(151668)
83 except ValueError:
84 index = 0
85
86 thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
87 content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
88
89 print("thinking content:", thinking_content)
90 print("content:", content)
91 ```
92
93 For deployment, you can use `sglang>=0.4.6.post1` or `vllm>=0.8.5` or to create an OpenAI-compatible API endpoint:
94 - SGLang:
95 ```shell
96 python -m sglang.launch_server --model-path Qwen/Qwen3-32B --reasoning-parser qwen3
97 ```
98 - vLLM:
99 ```shell
100 vllm serve Qwen/Qwen3-32B --enable-reasoning --reasoning-parser deepseek_r1
101 ```
102
103 For local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3.
104
105 ## Switching Between Thinking and Non-Thinking Mode
106
107 > [!TIP]
108 > The `enable_thinking` switch is also available in APIs created by SGLang and vLLM.
109 > 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.
110
111 ### `enable_thinking=True`
112
113 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.
114
115 ```python
116 text = tokenizer.apply_chat_template(
117 messages,
118 tokenize=False,
119 add_generation_prompt=True,
120 enable_thinking=True # True is the default value for enable_thinking
121 )
122 ```
123
124 In this mode, the model will generate think content wrapped in a `<think>...</think>` block, followed by the final response.
125
126 > [!NOTE]
127 > 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.
128
129
130 ### `enable_thinking=False`
131
132 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.
133
134 ```python
135 text = tokenizer.apply_chat_template(
136 messages,
137 tokenize=False,
138 add_generation_prompt=True,
139 enable_thinking=False # Setting enable_thinking=False disables thinking mode
140 )
141 ```
142
143 In this mode, the model will not generate any think content and will not include a `<think>...</think>` block.
144
145 > [!NOTE]
146 > 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.
147
148 ### Advanced Usage: Switching Between Thinking and Non-Thinking Modes via User Input
149
150 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.
151
152 Here is an example of a multi-turn conversation:
153
154 ```python
155 from transformers import AutoModelForCausalLM, AutoTokenizer
156
157 class QwenChatbot:
158 def __init__(self, model_name="Qwen/Qwen3-32B"):
159 self.tokenizer = AutoTokenizer.from_pretrained(model_name)
160 self.model = AutoModelForCausalLM.from_pretrained(model_name)
161 self.history = []
162
163 def generate_response(self, user_input):
164 messages = self.history + [{"role": "user", "content": user_input}]
165
166 text = self.tokenizer.apply_chat_template(
167 messages,
168 tokenize=False,
169 add_generation_prompt=True
170 )
171
172 inputs = self.tokenizer(text, return_tensors="pt")
173 response_ids = self.model.generate(**inputs, max_new_tokens=32768)[0][len(inputs.input_ids[0]):].tolist()
174 response = self.tokenizer.decode(response_ids, skip_special_tokens=True)
175
176 # Update history
177 self.history.append({"role": "user", "content": user_input})
178 self.history.append({"role": "assistant", "content": response})
179
180 return response
181
182 # Example Usage
183 if __name__ == "__main__":
184 chatbot = QwenChatbot()
185
186 # First input (without /think or /no_think tags, thinking mode is enabled by default)
187 user_input_1 = "How many r's in strawberries?"
188 print(f"User: {user_input_1}")
189 response_1 = chatbot.generate_response(user_input_1)
190 print(f"Bot: {response_1}")
191 print("----------------------")
192
193 # Second input with /no_think
194 user_input_2 = "Then, how many r's in blueberries? /no_think"
195 print(f"User: {user_input_2}")
196 response_2 = chatbot.generate_response(user_input_2)
197 print(f"Bot: {response_2}")
198 print("----------------------")
199
200 # Third input with /think
201 user_input_3 = "Really? /think"
202 print(f"User: {user_input_3}")
203 response_3 = chatbot.generate_response(user_input_3)
204 print(f"Bot: {response_3}")
205 ```
206
207 > [!NOTE]
208 > 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.
209 > 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.
210
211 ## Agentic Use
212
213 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.
214
215 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.
216 ```python
217 from qwen_agent.agents import Assistant
218
219 # Define LLM
220 llm_cfg = {
221 'model': 'Qwen3-32B',
222
223 # Use the endpoint provided by Alibaba Model Studio:
224 # 'model_type': 'qwen_dashscope',
225 # 'api_key': os.getenv('DASHSCOPE_API_KEY'),
226
227 # Use a custom endpoint compatible with OpenAI API:
228 'model_server': 'http://localhost:8000/v1', # api_base
229 'api_key': 'EMPTY',
230
231 # Other parameters:
232 # 'generate_cfg': {
233 # # Add: When the response content is `<think>this is the thought</think>this is the answer;
234 # # Do not add: When the response has been separated by reasoning_content and content.
235 # 'thought_in_content': True,
236 # },
237 }
238
239 # Define Tools
240 tools = [
241 {'mcpServers': { # You can specify the MCP configuration file
242 'time': {
243 'command': 'uvx',
244 'args': ['mcp-server-time', '--local-timezone=Asia/Shanghai']
245 },
246 "fetch": {
247 "command": "uvx",
248 "args": ["mcp-server-fetch"]
249 }
250 }
251 },
252 'code_interpreter', # Built-in tools
253 ]
254
255 # Define Agent
256 bot = Assistant(llm=llm_cfg, function_list=tools)
257
258 # Streaming generation
259 messages = [{'role': 'user', 'content': 'https://qwenlm.github.io/blog/ Introduce the latest developments of Qwen'}]
260 for responses in bot.run(messages=messages):
261 pass
262 print(responses)
263 ```
264
265 ## Processing Long Texts
266
267 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.
268
269 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:
270
271 - Modifying the model files:
272 In the `config.json` file, add the `rope_scaling` fields:
273 ```json
274 {
275 ...,
276 "rope_scaling": {
277 "rope_type": "yarn",
278 "factor": 4.0,
279 "original_max_position_embeddings": 32768
280 }
281 }
282 ```
283 For `llama.cpp`, you need to regenerate the GGUF file after the modification.
284
285 - Passing command line arguments:
286
287 For `vllm`, you can use
288 ```shell
289 vllm serve ... --rope-scaling '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' --max-model-len 131072
290 ```
291
292 For `sglang`, you can use
293 ```shell
294 python -m sglang.launch_server ... --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":32768}}'
295 ```
296
297 For `llama-server` from `llama.cpp`, you can use
298 ```shell
299 llama-server ... --rope-scaling yarn --rope-scale 4 --yarn-orig-ctx 32768
300 ```
301
302 > [!IMPORTANT]
303 > If you encounter the following warning
304 > ```
305 > Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'original_max_position_embeddings'}
306 > ```
307 > please upgrade `transformers>=4.51.0`.
308
309 > [!NOTE]
310 > 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.**
311 > We advise adding the `rope_scaling` configuration only when processing long contexts is required.
312 > 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.
313
314 > [!NOTE]
315 > 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.
316
317 > [!TIP]
318 > The endpoint provided by Alibaba Model Studio supports dynamic YaRN by default and no extra configuration is needed.
319
320 ## Best Practices
321
322 To achieve optimal performance, we recommend the following settings:
323
324 1. **Sampling Parameters**:
325 - 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.
326 - For non-thinking mode (`enable_thinking=False`), we suggest using `Temperature=0.7`, `TopP=0.8`, `TopK=20`, and `MinP=0`.
327 - 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.
328
329 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.
330
331 3. **Standardize Output Format**: We recommend using prompts to standardize model outputs when benchmarking.
332 - **Math Problems**: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
333 - **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"`."
334
335 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.
336
337 ### Citation
338
339 If you find our work helpful, feel free to give us a cite.
340
341 ```
342 @misc{qwen3technicalreport,
343 title={Qwen3 Technical Report},
344 author={Qwen Team},
345 year={2025},
346 eprint={2505.09388},
347 archivePrefix={arXiv},
348 primaryClass={cs.CL},
349 url={https://arxiv.org/abs/2505.09388},
350 }
351 ```