Qwen3是什么？2025混合思维AI大模型详解

引言

2025年伊始，我们隆重推出Qwen3系列模型，其旗舰型号Qwen3-235B-A22B现已正式可用。Qwen3标志着大型语言模型发展的一个重要里程碑，它引入了创新的混合思维AI范式，结合了强大的推理能力与前所未有的效率。该系列模型支持119种语言，并采用了革命性的MoE（专家混合）架构，旨在为复杂任务提供深度解决方案，同时保持卓越的资源效率。

At the beginning of 2025, we are proud to announce the official availability of the Qwen3 model series, with its flagship model, Qwen3-235B-A22B, now ready for use. Qwen3 represents a significant milestone in the evolution of large language models, introducing an innovative hybrid-thinking AI paradigm that combines powerful reasoning capabilities with unprecedented efficiency. Supporting 119 languages and built on a revolutionary Mixture-of-Experts (MoE) architecture, this series is designed to deliver in-depth solutions for complex tasks while maintaining exceptional resource efficiency.

什么是Qwen3？

Qwen3是我们最新一代的大型语言模型家族。其核心在于MoE架构混合专家（Mixture of Experts）架构，用于大规模模型训练，在阿里云平台中实现万卡规模训练，MFU达35%-40%。，该架构通过将先进的推理能力与高效的处理机制相结合，正在重塑人工智能的格局。基于高达36万亿token的训练数据，Qwen3在代码生成、数学解题、逻辑推理以及多语言任务处理等方面均展现出卓越的性能。

Qwen3 is our latest family of large language models. At its heart lies the Mixture-of-Experts (MoE) architecture, which is reshaping the AI landscape by combining advanced reasoning with efficient processing. Trained on an extensive corpus of 36 trillion tokens, Qwen3 delivers exceptional performance across a wide range of domains, including coding, mathematics, reasoning, and multilingual tasks.

Qwen3的核心特性

混合思维模式Qwen3的核心特性，允许AI在深度推理和快速响应之间动态切换，根据任务复杂度配置思维预算。

Qwen3允许用户在深度推理模式和快速响应模式之间灵活切换。对于复杂的逻辑或数学问题，可以启用深度推理以获得逐步、严谨的解决方案；而对于常规的对话或信息查询，则可使用快速响应模式以提升效率。用户还可以通过配置“思维预算”来精细控制模型在推理上投入的计算资源。

Qwen3 enables flexible switching between in-depth reasoning and quick-response modes. For complex logical or mathematical problems, users can activate deep reasoning to obtain step-by-step, rigorous solutions. For routine conversations or information queries, the quick-response mode offers greater efficiency. Additionally, users can finely control the computational resources allocated to reasoning by configuring a "thinking budget."

专家混合架构

MoE是一种革命性的模型架构。与传统稠密模型不同，MoE模型在每一层中包含了多个“专家”网络。在处理每个输入token时，系统仅激活最相关的少数几个专家，而非全部参数。这种设计在保持甚至超越大型稠密模型性能的同时，大幅降低了训练和推理过程中的计算成本与延迟。

Mixture-of-Experts (MoE) is a revolutionary model architecture. Unlike traditional dense models, MoE models incorporate multiple "expert" networks within each layer. For each input token, the system activates only the most relevant subset of experts, rather than all parameters. This design significantly reduces computational costs and latency during both training and inference, while maintaining or even surpassing the performance of large dense models.

广泛的多语言支持

Qwen3具备强大的多语言理解和生成能力，覆盖全球119种语言和方言，从西欧语言到南亚语言均在其支持范围内。这使得Qwen3能够无缝应用于全球化产品、跨语言研究以及多语言内容创作等场景。

Qwen3 possesses robust multilingual understanding and generation capabilities, covering 119 global languages and dialects, ranging from Western European to Southern Asian languages. This enables Qwen3 to be seamlessly applied in scenarios such as globalized products, cross-lingual research, and multilingual content creation.

扩展的上下文长度

模型支持长达128K token的上下文窗口，能够一次性处理和分析超长文档、代码库或复杂的多轮对话，确保信息处理的连贯性和完整性，无“盲点”。

The model supports an extended context window of up to 128K tokens, enabling it to process and analyze lengthy documents, codebases, or complex multi-turn conversations in a single pass. This ensures coherence and completeness in information handling, with no "blind spots."

为何选择Qwen3？

Qwen3集尖端能力于一身，旨在高效处理复杂任务，代表了AI的未来发展方向。在多项关键基准测试中，例如AIME（国际数学竞赛）、LiveCodeBench（实时代码生成评测）和BFCL（基础能力评测）中，Qwen3的表现与DeepSeek-R1、o1、o3-mini、Gemini-2.5-Pro等顶尖模型相比极具竞争力。

Qwen3 integrates cutting-edge capabilities designed to handle complex tasks efficiently, representing the future direction of AI. In several key benchmarks, such as AIME (American Invitational Mathematics Examination), LiveCodeBench, and BFCL (Basic Functionality and Capability Leaderboard), Qwen3 demonstrates highly competitive performance compared to top-tier models like DeepSeek-R1, o1, o3-mini, and Gemini-2.5-Pro.

其优势具体体现在：

卓越的基准测试成绩：在Arena-Hard、LiveBench、GPQA-Diamond、MMLU-Pro等评测中表现优异。
强大的模型家族：提供从0.6B到235B参数的八种规格模型，包含两个高效的MoE模型，满足不同场景需求。
便捷的AI就绪部署：预配置支持SGLang、vLLM等推理框架，并可轻松创建与OpenAI兼容的API端点。
开放的开源许可：所有模型均在Apache 2.0许可证一种宽松的开源软件许可证，允许用户自由使用、修改和分发软件，包括商业用途，仅需保留版权声明。下发布，允许商业和非商业用途，为研究和应用提供极大灵活性。

Its advantages are specifically reflected in:

Excellent Benchmark Performance: Outstanding results in evaluations such as Arena-Hard, LiveBench, GPQA-Diamond, and MMLU-Pro.

Robust Model Family: Offers eight model sizes ranging from 0.6B to 235B parameters, including two efficient MoE models to meet diverse scenario requirements.

AI-Ready Deployment: Pre-configured for easy deployment with inference frameworks like SGLang and vLLM, and can easily create OpenAI-compatible API endpoints.

Open-Source License: All models are released under the permissive Apache 2.0 license, allowing for both commercial and non-commercial use, providing great flexibility for research and application.

用户评价

来自全球的开发者和研究人员分享了他们使用Qwen3的体验：

Developers and researchers from around the world have shared their experiences using Qwen3:

David Chen, AI研究员
“Qwen3的混合思维模式Qwen3的核心特性，允许AI在深度推理和快速响应之间动态切换，根据任务复杂度配置思维预算。彻底改变了我们的研究工作流。通过配置思维预算，我们前所未有地掌控了性能与效率之间的平衡。”

David Chen, AI Researcher
“Qwen3's hybrid thinking modes have revolutionized our research workflow. The ability to configure thinking budgets gives us unprecedented control over the balance between performance and efficiency.”

Rachel Kim, 软件开发者
“Qwen3的MoE架构混合专家（Mixture of Experts）架构，用于大规模模型训练，在阿里云平台中实现万卡规模训练，MFU达35%-40%。以极低的计算成本提供了卓越的性能。我们得以部署以往对我们机构来说遥不可及的复杂AI能力。”

Rachel Kim, Software Developer
“The MoE architecture in Qwen3 delivers exceptional performance at a fraction of the computational cost. We've been able to deploy sophisticated AI capabilities that were previously out of reach for our organization.”

Michael Johnson, NLP工程师
“Qwen3对119种语言的支持是我们多语言应用的改变者。该模型以惊人的准确性处理复杂的翻译任务和跨语言理解。”

Michael Johnson, NLP Engineer
“Qwen3's support for 119 languages has been a game-changer for our multilingual applications. The model handles complex translation tasks and cross-lingual understanding with remarkable accuracy.”

常见问题解答

1. Qwen3与其他大型语言模型有何不同？
Qwen3引入了混合思维模式Qwen3的核心特性，允许AI在深度推理和快速响应之间动态切换，根据任务复杂度配置思维预算。，允许模型在深度推理和快速响应之间切换。结合MoE架构混合专家（Mixture of Experts）架构，用于大规模模型训练，在阿里云平台中实现万卡规模训练，MFU达35%-40%。，它在降低计算需求的同时提供了卓越性能。此外，模型支持119种语言，并具备长达128K token的上下文长度。

1. What makes Qwen3 different from other large language models?
Qwen3 introduces hybrid thinking modes, allowing the model to switch between deep reasoning and quick responses. Combined with the MoE architecture, this delivers exceptional performance with lower computational requirements. Additionally, the models support 119 languages and feature an extended context length of up to 128K tokens.

2. 如何控制Qwen3的思维模式？
您可以通过调用模型时的 enable_thinking 参数进行控制。设置为 True 启用深度推理，False 则提供快速响应。在多轮对话中，您还可以在提示词中使用 /think 和 /no_think 命令来动态切换模式。

2. How can I control the thinking modes in Qwen3?
You can control Qwen3's thinking modes via the enable_thinking parameter when calling the model. Setting it to True enables in-depth reasoning, while False provides quicker responses. During multi-turn conversations, you can also use the /think and /no_think commands within prompts to dynamically switch between modes.

3. 可以使用Qwen3构建哪些类型的应用？
Qwen3支持广泛的AI应用，从内容生成到复杂推理任务。该模型在编码、数学、逻辑推理和多语言翻译方面表现出色，适用于构建聊天机器人、研究助手、创意写作工具等。

3. What types of applications can I build with Qwen3?
Qwen3 supports a wide range of AI applications, from content generation to complex reasoning tasks. The models excel at coding, mathematics, logical reasoning, and multilingual translation, making them suitable for building chatbots, research assistants, creative writing tools, and more.

4. Qwen3有哪些部署选项？
Qwen3模型可以使用SGLang、vLLM等框架进行部署，以创建与OpenAI兼容的API端点。对于本地使用，可以使用Ollama、LM Studio、MLX、llama.cpp或KTransformers等工具。所有模型均可从Hugging Face、ModelScope和Kaggle平台下载，遵循Apache 2.0许可证一种宽松的开源软件许可证，允许用户自由使用、修改和分发软件，包括商业用途，仅需保留版权声明。。

4. What deployment options are available for Qwen3?
Qwen3 models can be deployed using frameworks like SGLang and vLLM to create OpenAI-compatible API endpoints. For local usage, tools such as Ollama, LM Studio, MLX, llama.cpp, or KTransformers can be used. All models are available for download from Hugging Face, ModelScope, and Kaggle under the Apache 2.0 license.

准备好体验Qwen3的强大功能了吗？

立即开始使用我们最先进的大型语言模型，开启您的AI创新之旅。

Ready to experience the power of Qwen3? Start building with our state-of-the-art large language models today and embark on your journey of AI innovation.