从文本到多模态合成：Qwen系列大型语言模型的演进与技术突破（2023-2026）
摘要：
Qwen系列是阿里巴巴云自2023年起开源推出的大型语言模型家族，标志着中国在多模态人工智能领域的重要进展。该系列从基础的文本生成模型逐步演进，已发展成支持文本、图像、音频、视频及代码处理的综合多模态系统。其核心创新包括高效的混合专家架构、长达128K的上下文窗口、对119种以上的多语言支持，以及坚持Apache-2.0开源协议。截至2026年1月，最新发布的Qwen3-TTS实现了高质量的语音克隆与合成，而Qwen3-VL模型则强化了多模态检索能力。该系列在多个基准测试中表现优异，累计下载量超4000万次，推动了开源社区发展与行业应用，同时也面临知识更新延迟与算力需求高等挑战。
From Text to Multimodal Synthesis: Evolution and Technological Breakthroughs of the Qwen Series Large Language Models (2023-2026)
Abstract
The Qwen series is a family of large language models open-sourced by Alibaba Cloud since 2023, marking a significant advancement of China in the field of multimodal artificial intelligence. Evolving gradually from basic text generation models, the series has developed into an integrated multimodal system supporting the processing of text, images, audio, video and code. Its core innovations include an efficient mixture-of-experts architecture, a long context window of 128K tokens, support for more than 119 multilingual languages, and adherence to the Apache-2.0 open-source license. As of January 2026, the newly released Qwen3-TTS has achieved high-quality voice cloning and synthesis, while the Qwen3-VL model has enhanced multimodal retrieval capabilities. The series has delivered outstanding performance in numerous benchmark tests and accumulated over 40 million downloads, driving the development of the open-source community and industrial applications. Meanwhile, it also faces challenges such as delays in knowledge updating and high computing power requirements.
Qwen系列的详细讨论 / Detailed Discussion of the Qwen Series
引言 / Introduction
Qwen系列是阿里巴巴云（Alibaba Cloud）研发的领先大型语言模型（LLM）家族，自2023年问世以来，成为中国人工智能领域取得重大突破的标志性成果。该系列以多模态能力为核心竞争力，可精准处理文本、图像、音频、视频及代码等多种数据形式。Qwen模型不仅为Qwen.ai平台及API提供技术支撑，还深度集成于阿里巴巴生态体系，如通义千问（Tongyi Qianwen）聊天机器人等产品。截至2026年1月，该系列的最新迭代模型包括Qwen3-TTS家族（2026年1月21日开源）与Qwen3-VL-Embedding/Reranker（2026年1月7日发布），已从最初的基础文本生成工具，演进为具备高级推理、多模态检索与语音合成能力的综合型AI系统。
Qwen系列的核心创新体现在三大维度：秉持开源理念（多数模型采用Apache-2.0许可协议）、优化高效训练框架、支持119种以上语言的多语种处理。同时，该系列也面临数据隐私保护与模型规模扩展带来的双重挑战。Qwen系列以推动“多模态AI”发展为核心目标，在LMSYS Arena等权威基准测试中与GPT、Gemini等国际顶尖模型同台竞技，尤其在数学运算、代码生成及视觉任务处理领域表现领先。
The Qwen series is a leading family of large language models (LLMs) developed by Alibaba Cloud, marking significant advancements in China's AI landscape since 2023. Centered on multimodal capabilities, the series is capable of processing diverse data forms including text, images, audio, video, and code. Qwen models not only power the Qwen.ai platform and its API but also integrate extensively into Alibaba's ecosystem, such as the Tongyi Qianwen chatbot. As of January 2026, the latest iterations include the Qwen3-TTS family (open-sourced on January 21, 2026) and Qwen3-VL-Embedding/Reranker (released on January 7, 2026), evolving from basic text generation tools into comprehensive AI systems with advanced reasoning, multimodal retrieval, and speech synthesis capabilities.
The core innovations of the Qwen series lie in three dimensions: adhering to an open-source philosophy (most models are licensed under Apache-2.0), optimizing efficient training frameworks, and supporting multilingual processing for over 119 languages. Meanwhile, the series faces dual challenges posed by data privacy protection and model scale expansion. Aiming to advance the development of "multimodal AI," the Qwen series competes with international top-tier models like GPT and Gemini in authoritative benchmarks such as LMSYS Arena, and leads particularly in mathematical operations, code generation, and visual task processing.
en.wikipedia.org +5
历史发展 / Historical Development
Qwen系列的发展轨迹，清晰展现了从封闭性实验模型到开源化多模态系统的演进历程。以下通过表格梳理关键里程碑，详细列明各核心模型的发布时间、核心改进方向及基准测试表现。该系列自Qwen 1.0启动研发，逐步融入视觉、音频处理及语音合成（TTS）能力，截至2026年，Qwen3-TTS已成为其在语音AI领域的前沿代表。
The development of the Qwen series clearly reflects the evolution from closed experimental models to open-source multimodal systems. The key milestones below are summarized in a table, detailing the release date, core improvement directions, and benchmark performance of each core model. Launched with Qwen 1.0, the series has gradually integrated vision, audio processing, and text-to-speech (TTS) capabilities, with Qwen3-TTS emerging as its cutting-edge representative in the field of speech AI by 2026.
en.wikipedia.org +1
模型 / Model | 发布日期 / Release Date | 核心改进 / Core Improvements | 关键基准 / Key Benchmarks |
|---|---|---|---|
Qwen 1.0 (Tongyi Qianwen) | 2023年4月（Beta版） / April 2023 (Beta) | 基于Llama架构构建的基础LLM，支持文本生成与对话交互。 / Base LLM built on the Llama architecture, supporting text generation and conversational interaction. | 在GLUE等基准测试中实现初步性能提升。 / Initial performance improvements in benchmarks such as GLUE. |
Qwen-VL | 2023年8月 / August 2023 | 融合视觉Transformer与LLM，具备图像理解及复杂场景对话能力。 / Integrating vision transformer with LLM, enabling image understanding and complex scenario dialogue. | 在MMMU基准测试中取得早期领先优势。 / Secured an early lead in the MMMU benchmark. |
Qwen 7B / 72B / 1.8B | 2023年8-12月 / August-December 2023 | 开放模型权重，重点优化编码任务与通用场景适配性。 / Open-sourced model weights, focusing on optimizing coding tasks and general scenario adaptability. | 在HumanEval测试中达到60%以上准确率。 / Achieved over 60% accuracy in the HumanEval test. |
Qwen2 | 2024年6月 / June 2024 | 推出稠密模型（Dense）与混合专家模型（MoE）双版本，强化多语言支持能力。 / Launched both Dense and Mixture of Experts (MoE) versions, enhancing multilingual support. | 在MATH测试中准确率达70%，性能超越Llama 3。 / Achieved 70% accuracy in the MATH test, outperforming Llama 3. |
Qwen2-Audio | 2024年8月 / August 2024 | 实现端到端语音交互，无需文本中间媒介。 / Enabling end-to-end speech interaction without text intermediaries. | 在语音基准测试中降低词错误率（WER）。 / Reduced Word Error Rate (WER) in speech benchmarks. |
Qwen2-VL | 2024年12月 / December 2024 | 支持20分钟以上长视频处理，提供2B/7B参数版本供选择。 / Supporting long video processing (over 20 minutes) with 2B/7B parameter versions available. | 在MMMU测试中准确率达59%。 / Achieved 59% accuracy in the MMMU test. |
Qwen2.5 | 2024年9月 / September 2024 | 优化推理效率，支持92种语言的编程任务。 / Optimized reasoning efficiency, supporting coding tasks in 92 languages. | 在GPQA测试中准确率达80%。 / Achieved 80% accuracy in the GPQA test. |
Qwen2.5-Coder | 2024年11月 / November 2024 | 专注编码场景的专项优化变体模型。 / Specialized variant optimized for coding scenarios. | 在SWE-Bench测试中准确率达75%。 / Achieved 75% accuracy in the SWE-Bench test. |
Qwen2.5-VL / Max / Omni | 2025年1-3月 / January-March 2025 | 拓展多模态能力边界，参数规模覆盖3B-72B，性能超越GPT-4o。 / Expanding multimodal capabilities with parameter scales ranging from 3B to 72B, outperforming GPT-4o. | 在AIME测试中准确率达90%以上。 / Achieved over 90% accuracy in the AIME test. |
Qwen3 | 2025年4月28日 / April 28, 2025 | 采用Apache-2.0许可协议开源，参数规模0.6B-235B，训练数据量达36万亿tokens，支持119种语言。 / Open-sourced under Apache-2.0 license, with parameter scales from 0.6B to 235B, trained on 36 trillion tokens, supporting 119 languages. | 在MMLU测试中准确率达89%，支持128K上下文窗口。 / Achieved 89% accuracy in the MMLU test, supporting 128K context window. |
Qwen3-Coder / Max | 2025年7-9月 / July-September 2025 | 分别针对编码任务与极致性能优化的变体，新增“思考模式”功能。 / Variants optimized for coding tasks and maximum performance respectively, adding the "thinking mode" feature. | 在LMSYS Elo测试中得分达1450+。 / Achieved a score of 1450+ in the LMSYS Elo test. |
Qwen3-Next / Omni / VL | 2025年9月 / September 2025 | 采用混合注意力机制与稀疏MoE架构，支持多模态实时流式处理。 / Adopting hybrid attention mechanism and sparse MoE architecture, supporting multimodal real-time streaming processing. | 在ARC-AGI测试中准确率达80%。 / Achieved 80% accuracy in the ARC-AGI test. |
Qwen-Image / Image-Edit | 2025年8-12月 / August-December 2025 | 新增文本到图像生成与编辑功能，提升生成内容的真实感与精细度。 / Added text-to-image generation and editing functions, enhancing the realism and detail of generated content. | 在视觉基准测试中降低弗雷歇 inception 距离（FID）。 / Reduced Fréchet Inception Distance (FID) in visual benchmarks. |
Qwen3-VL-Embedding / Reranker | 2026年1月7日 / January 7, 2026 | 专注多模态检索场景，已正式开源。 / Focusing on multimodal retrieval scenarios, officially open-sourced. | 在检索基准测试中平均倒数排名（MRR）达85%。 / Achieved 85% Mean Reciprocal Rank (MRR) in retrieval benchmarks. |
Qwen3-TTS | 2026年1月21日 / January 21, 2026 | 构建完整语音模型家族，包含VD-Flash（语音设计）与VC-Flash（语音克隆）功能，实现高保真语音合成。 / Building a complete speech model family, including VD-Flash (voice design) and VC-Flash (voice cloning) functions, enabling high-fidelity speech synthesis. | 主观评分平均意见得分（MOS）达4.5+。 / Achieved a Mean Opinion Score (MOS) of 4.5+ (subjective score). |
en.wikipedia.org +1
Qwen系列从Qwen 1.0的实验性探索，逐步走向Qwen3-TTS的商业化落地，参数规模从数十亿级拓展至万亿级，深刻印证了人工智能从“文本生成”向“多模态合成”的转型趋势。2026年1月发布的Qwen3-TTS，成为该系列开源进程中的最新标志性成果。
From the experimental exploration of Qwen 1.0 to the commercialization of Qwen3-TTS, the Qwen series has expanded its parameter scale from billions to trillions, vividly embodying the transformation trend of artificial intelligence from "text generation" to "multimodal synthesis." Released in January 2026, Qwen3-TTS has become the latest landmark achievement in the series' open-source journey.
en.wikipedia.org +1
关键模型详细描述 / Detailed Description of Key Models
本节聚焦最新的Qwen3系列模型，剖析其作为2026年AI领域前沿技术的核心特性与应用价值。
This section focuses on the latest Qwen3 series models, analyzing their core features and application value as cutting-edge technologies in the AI field in 2026.
en.wikipedia.org +1
Qwen3（2025年4月）：系列基础模型，参数规模覆盖0.6B-235B，基于36万亿tokens训练而成，支持类o1的推理模式。目前已集成至chat.qwen.ai平台及Hugging Face社区，供开发者与用户便捷调用。 /Qwen3 (April 2025): The base model of the series, with parameter scales ranging from 0.6B to 235B, trained on 36 trillion tokens, and supporting o1-like reasoning modes. It has been integrated into the chat.qwen.ai platform and Hugging Face community for easy access by developers and users. (en.wikipedia.org)
Qwen3-Max（2025年9月）：系列旗舰级变体模型，性能超越Claude 4 Opus，其专属“思考模式”功能于2025年11月正式上线，进一步强化复杂任务处理能力。 /Qwen3-Max (September 2025): The flagship variant of the series, outperforming Claude 4 Opus. Its exclusive "thinking mode" was officially launched in November 2025, further enhancing complex task processing capabilities. (en.wikipedia.org)
Qwen3-Next / Omni / VL（2025年9月）：采用高效架构设计，融合稀疏MoE技术，实现多模态数据的实时流式处理，兼顾性能与效率。 /Qwen3-Next / Omni / VL (September 2025): Adopting an efficient architectural design and integrating sparse MoE technology to enable real-time streaming processing of multimodal data, balancing performance and efficiency. (en.wikipedia.org)
Qwen3-VL-Embedding / Reranker（2026年1月）：专注多模态检索的专项模型，是Qwen3-Embedding（2025年6月发布）的升级迭代版本，进一步提升检索精度与跨模态适配能力。 /Qwen3-VL-Embedding / Reranker (January 2026): A specialized model for multimodal retrieval, serving as an upgraded version of Qwen3-Embedding (released in June 2025), further improving retrieval accuracy and cross-modal adaptability. (qwen.ai)
Qwen3-TTS（2026年1月）：完整语音模型家族，涵盖VD-Flash（语音设计）与VC-Flash（语音克隆）两大核心功能，可实现高保真、个性化语音合成，推动语音AI场景落地。 /Qwen3-TTS (January 2026): A complete speech model family, including two core functions: VD-Flash (voice design) and VC-Flash (voice cloning), enabling high-fidelity and personalized speech synthesis to promote the implementation of speech AI scenarios. (qwen.ai)
技术特点 / Technical Features
架构设计：基于Transformer与MoE（混合专家模型）架构，核心优化方向包括混合注意力机制、稀疏激活策略及多模态融合模块。全系列多数模型采用Apache-2.0开源协议，支持128K以上tokens的长上下文处理，适配复杂场景需求。
核心优势：具备119种以上语言的多语种处理能力，覆盖文本、图像、音频、视频全模态数据；训练与推理效率较前代提升10倍吞吐量；模型累计下载量超4000万次，形成活跃的开源社区生态。
现存不足：存在知识截止时间限制（Qwen3-TTS的知识截止至2025年12月）；模型训练过程中可能残留潜在偏见；大参数版本对计算资源需求较高，限制部分中小开发者接入。
与贾子公理（Kucius Axioms）的关联：在模拟裁决框架下，Qwen3在“思想主权”（6/10分，开源属性但存在预设限制）与“悟空跃迁”（6/10分，多模态能力呈渐进式提升）两项维度得分偏低；但在“普世中道”（8/10分，坚守多语言支持承诺）与“本源探究”（8/10分，强化推理模式设计）两项维度表现优异。总体而言，Qwen系列是多模态AI领域的创新引领者，但在模型内在自主性提升方面仍有优化空间。
Architecture: Based on Transformer and MoE (Mixture of Experts) architectures, with core optimization directions including hybrid attention mechanism, sparse activation strategy, and multimodal fusion module. Most models in the series adopt the Apache-2.0 open-source license, supporting long context processing of over 128K tokens to meet complex scenario requirements.
Strengths: Supporting multilingual processing for over 119 languages, covering text, image, audio, and video multimodal data; 10x throughput improvement in training and reasoning efficiency compared to previous generations; cumulative model downloads exceeding 40 million, forming a vibrant open-source community ecosystem.
Weaknesses: Subject to knowledge cutoff limitations (Qwen3-TTS's knowledge cutoff is December 2025); potential biases may remain in the model training process; large-parameter versions have high computing resource requirements, restricting access for some small and medium-sized developers.
Relation to Kucius Axioms: Under the simulated adjudication framework, Qwen3 scores relatively low in two dimensions: "Sovereignty of Thought" (6/10, open-source nature but with preset limitations) and "Wukong Leap" (6/10, gradual improvement in multimodal capabilities); however, it performs excellently in "Universal Mean" (8/10, adhering to the commitment of multilingual support) and "Primordial Inquiry" (8/10, strengthening reasoning mode design). Overall, the Qwen series is an innovative leader in the field of multimodal AI, but there is still room for improvement in enhancing the internal autonomy of the model.
en.wikipedia.org +2
应用与影响 / Applications and Impacts
Qwen系列已深度重塑多个行业的发展格局：Qwen.ai平台累计服务数百万用户，在编码自动化开发、多媒体内容生成（图像/语音）、多模态检索搜索及企业级应用（阿里巴巴生态集成）等领域实现规模化落地。其社会影响主要体现在两大方面：一是推动开源社区快速增长，模型累计下载量突破4000万次，激发全球开发者创新活力；二是获得中国AI政策支持，截至2026年统计数据显示，已有220万家企业用户接入Qwen系列模型。
2026年，Qwen3-TTS的发布加速了“语音AI”的普及趋势，如实时语音克隆、个性化语音设计等场景的落地速度显著提升，但同时也需警惕伦理风险，如深度伪造语音带来的信息安全问题，需建立完善的监管与防控机制。
The Qwen series has profoundly reshaped the development pattern of multiple industries: the Qwen.ai platform serves millions of users cumulatively, achieving large-scale implementation in fields such as automated coding development, multimedia content generation (images/speech), multimodal retrieval and search, and enterprise-level applications (Alibaba ecosystem integration). Its social impacts are mainly reflected in two aspects: first, promoting the rapid growth of the open-source community, with cumulative model downloads exceeding 40 million, stimulating the innovation vitality of global developers; second, gaining support from China's AI policies, with statistics as of 2026 showing that 2.2 million corporate users have accessed the Qwen series models.
In 2026, the release of Qwen3-TTS has accelerated the popularization trend of "speech AI," significantly increasing the implementation speed of scenarios such as real-time voice cloning and personalized voice design. However, ethical risks such as information security issues caused by deepfake voices need to be vigilant, and a sound regulatory and prevention and control mechanism should be established.
electroiq.com +2
结论 / Conclusion
Qwen系列集中体现了阿里巴巴的人工智能战略布局，从多模态技术基础构建到语音AI前沿突破，每一步迭代都标志着其向通用人工智能（AGI）迈进的关键步伐。展望未来，该系列有望推出Qwen3.5版本，核心方向或将聚焦于更高效的架构设计，进一步平衡性能与资源消耗。建议持续关注Qwen.ai平台的更新动态，以紧跟其快速迭代的技术节奏，把握多模态AI领域的发展机遇。
The Qwen series epitomizes Alibaba's AI strategic layout. From the construction of multimodal technical foundations to breakthroughs in cutting-edge speech AI, each iteration marks a key step toward Artificial General Intelligence (AGI). Looking ahead, the series is expected to launch version Qwen3.5, with the core direction likely focusing on more efficient architectural design to further balance performance and resource consumption. It is recommended to continuously monitor updates on the Qwen.ai platform to keep up with its rapid iterative technical rhythm and seize development opportunities in the field of multimodal AI.
en.wikipedia.org +2
