{"repo": "barada02/livemodel", "files": {"ModelsInfo/qwen_for_llm.md": {"6866e62e279bef74eb9ba48f3e24d41ed09b5b84": {"t": 1786068575.1515613, "n": 12673, "h": "ccf3f576efa87ae9", "c": "Qwen3-0.6B\nChat\nQwen3 Highlights\nQwen3 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:\n\nUniquely 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.\nSignificantly 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.\nSuperior 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.\nExpertise 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.\nSupport of 100+ languages and dialects with strong capabilities for multilingual instruction following and translation.\nModel Overview\nQwen3-0.6B has the following features:\n\nType: Causal Language Models\nTraining Stage: Pretraining & Post-training\nNumber of Parameters: 0.6B\nNumber of Paramaters (Non-Embedding): 0.44B\nNumber of Layers: 28\nNumber of Attention Heads (GQA): 16 for Q and 8 for KV\nContext Length: 32,768\nFor more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation.\n\nIf you encounter significant endless repetitions, please refer to the Best Practices section for optimal sampling parameters, and set the presence_penalty to 1.5.\n\nQuickstart\nThe code of Qwen3 has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers.\n\nWith transformers<4.51.0, you will encounter the following error:\n\nKeyError: 'qwen3'\n\nThe following contains a code snippet illustrating how to use the model generate content based on given inputs.\n\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel_name = \"Qwen/Qwen3-0.6B\"\n\n# load the tokenizer and the model\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_name,\n    torch_dtype=\"auto\",\n    device_map=\"auto\"\n)\n\n# prepare the model input\nprompt = \"Give me a short introduction to large language model.\"\nmessages = [\n    {\"role\": \"user\", \"content\": prompt}\n]\ntext = tokenizer.apply_chat_template(\n    messages,\n    tokenize=False,\n    add_generation_prompt=True,\n    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.\n)\nmodel_inputs = tokenizer([text], return_tensors=\"pt\").to(model.device)\n\n# conduct text completion\ngenerated_ids = model.generate(\n    **model_inputs,\n    max_new_tokens=32768\n)\noutput_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() \n\n# parsing thinking content\ntry:\n    # rindex finding 151668 (</think>)\n    index = len(output_ids) - output_ids[::-1].index(151668)\nexcept ValueError:\n    index = 0\n\nthinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip(\"\\n\")\ncontent = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip(\"\\n\")\n\nprint(\"thinking content:\", thinking_content)\nprint(\"content:\", content)\n\nFor deployment, you can use sglang>=0.4.6.post1 or vllm>=0.8.5 or to create an OpenAI-compatible API endpoint:\n\nSGLang:\npython -m sglang.launch_server --model-path Qwen/Qwen3-0.6B --reasoning-parser qwen3\n\nvLLM:\nvllm serve Qwen/Qwen3-0.6B --enable-reasoning --reasoning-parser deepseek_r1\n\nFor local use, applications such as Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers have also supported Qwen3.\n\nSwitching Between Thinking and Non-Thinking Mode\nThe enable_thinking switch is also available in APIs created by SGLang and vLLM. Please refer to our documentation for SGLang and vLLM users.\n\nenable_thinking=True\nBy 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.\n\ntext = tokenizer.apply_chat_template(\n    messages,\n    tokenize=False,\n    add_generation_prompt=True,\n    enable_thinking=True  # True is the default value for enable_thinking\n)\n\nIn this mode, the model will generate think content wrapped in a <think>...</think> block, followed by the final response.\n\nFor 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 section.\n\nenable_thinking=False\nWe 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.\n\ntext = tokenizer.apply_chat_template(\n    messages,\n    tokenize=False,\n    add_generation_prompt=True,\n    enable_thinking=False  # Setting enable_thinking=False disables thinking mode\n)\n\nIn this mode, the model will not generate any think content and will not include a <think>...</think> block.\n\nFor 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 section.\n\nAdvanced Usage: Switching Between Thinking and Non-Thinking Modes via User Input\nWe 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.\n\nHere is an example of a multi-turn conversation:\n\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nclass QwenChatbot:\n    def __init__(self, model_name=\"Qwen/Qwen3-0.6B\"):\n        self.tokenizer = AutoTokenizer.from_pretrained(model_name)\n        self.model = AutoModelForCausalLM.from_pretrained(model_name)\n        self.history = []\n\n    def generate_response(self, user_input):\n        messages = self.history + [{\"role\": \"user\", \"content\": user_input}]\n\n        text = self.tokenizer.apply_chat_template(\n            messages,\n            tokenize=False,\n            add_generation_prompt=True\n        )\n\n        inputs = self.tokenizer(text, return_tensors=\"pt\")\n        response_ids = self.model.generate(**inputs, max_new_tokens=32768)[0][len(inputs.input_ids[0]):].tolist()\n        response = self.tokenizer.decode(response_ids, skip_special_tokens=True)\n\n        # Update history\n        self.history.append({\"role\": \"user\", \"content\": user_input})\n        self.history.append({\"role\": \"assistant\", \"content\": response})\n\n        return response\n\n# Example Usage\nif __name__ == \"__main__\":\n    chatbot = QwenChatbot()\n\n    # First input (without /think or /no_think tags, thinking mode is enabled by default)\n    user_input_1 = \"How many r's in strawberries?\"\n    print(f\"User: {user_input_1}\")\n    response_1 = chatbot.generate_response(user_input_1)\n    print(f\"Bot: {response_1}\")\n    print(\"----------------------\")\n\n    # Second input with /no_think\n    user_input_2 = \"Then, how many r's in blueberries? /no_think\"\n    print(f\"User: {user_input_2}\")\n    response_2 = chatbot.generate_response(user_input_2)\n    print(f\"Bot: {response_2}\") \n    print(\"----------------------\")\n\n    # Third input with /think\n    user_input_3 = \"Really? /think\"\n    print(f\"User: {user_input_3}\")\n    response_3 = chatbot.generate_response(user_input_3)\n    print(f\"Bot: {response_3}\")\n\nFor 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. 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.\n\nAgentic Use\nQwen3 excels in tool calling capabilities. We recommend using 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.\n\nTo define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.\n\nfrom qwen_agent.agents import Assistant\n\n# Define LLM\nllm_cfg = {\n    'model': 'Qwen3-0.6B',\n\n    # Use the endpoint provided by Alibaba Model Studio:\n    # 'model_type': 'qwen_dashscope',\n    # 'api_key': os.getenv('DASHSCOPE_API_KEY'),\n\n    # Use a custom endpoint compatible with OpenAI API:\n    'model_server': 'http://localhost:8000/v1',  # api_base\n    'api_key': 'EMPTY',\n\n    # Other parameters:\n    # 'generate_cfg': {\n    #         # Add: When the response content is `<think>this is the thought</think>this is the answer;\n    #         # Do not add: When the response has been separated by reasoning_content and content.\n    #         'thought_in_content': True,\n    #     },\n}\n\n# Define Tools\ntools = [\n    {'mcpServers': {  # You can specify the MCP configuration file\n            'time': {\n                'command': 'uvx',\n                'args': ['mcp-server-time', '--local-timezone=Asia/Shanghai']\n            },\n            \"fetch\": {\n                \"command\": \"uvx\",\n                \"args\": [\"mcp-server-fetch\"]\n            }\n        }\n    },\n  'code_interpreter',  # Built-in tools\n]\n\n# Define Agent\nbot = Assistant(llm=llm_cfg, function_list=tools)\n\n# Streaming generation\nmessages = [{'role': 'user', 'content': 'https://qwenlm.github.io/blog/ Introduce the latest developments of Qwen'}]\nfor responses in bot.run(messages=messages):\n    pass\nprint(responses)\n\nBest Practices\nTo achieve optimal performance, we recommend the following settings:\n\nSampling Parameters:\n\nFor 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.\nFor non-thinking mode (enable_thinking=False), we suggest using Temperature=0.7, TopP=0.8, TopK=20, and MinP=0.\nFor 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.\nAdequate 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.\n\nStandardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.\n\nMath Problems: Include \"Please reason step by step, and put your final answer within \\boxed{}.\" in the prompt.\nMultiple-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\".\"\nNo 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.\n\nCitation\nIf you find our work helpful, feel free to give us a cite.\n\n@misc{qwen3technicalreport,\n      title={Qwen3 Technical Report}, \n      author={Qwen Team},\n      year={2025},\n      eprint={2505.09388},\n      archivePrefix={arXiv},\n      primaryClass={cs.CL},\n      url={https://arxiv.org/abs/2505.09388}, \n}"}}}}