Qwen

Qwen（Chinese: 通义千问； pinyin: Tōngyì Qiānwèn），also known internationally as Tongyi Qianwen, is a family of large language models (LLMs) and multimodal artificial intelligence models developed by Alibaba Cloud, the cloud computing subsidiary of the Chinese technology group Alibaba Group. First introduced in April 2023, the Qwen series has grown to encompass models of widely varying scales—ranging from sub-billion-parameter compact models to systems exceeding one trillion parameters—as well as specialized variants for vision, speech, coding, mathematics, and multi-step reasoning. In addition to being offered through Alibaba Cloud's commercial model services, many Qwen models have been released as open weights, making Qwen one of the most widely downloaded and re-used open-source model families in the global AI community.

Background

Alibaba Group began investing heavily in artificial intelligence research in the late 2010s, notably through DAMO Academy, its fundamental research arm, which established labs dedicated to machine learning, natural language processing, and speech technology. In the early 2020s, these efforts were consolidated under the "Tongyi" (通义， loosely "universal meaning") initiative, which produced the M6 large-scale pretrained model and other research systems.

The public launch of ChatGPT in late 2022 triggered intense competition in large language models worldwide, and particularly among Chinese technology firms. In April 2023, Alibaba announced Tongyi Qianwen at its Alibaba Cloud Summit, positioning the model both as a consumer-facing AI assistant and as the foundation of a "model-as-a-service" strategy in which enterprises could build applications on Alibaba Cloud's AI infrastructure. The name Qianwen (千问)， meaning "a thousand questions," reflects the idea of pursuing understanding through extensive inquiry, while "Tongyi" anchors the model within Alibaba's broader AI program.

History

2023: Launch and first open-source releases. Tongyi Qianwen was unveiled on 11 April 2023 and rolled out in beta through Alibaba's enterprise communication platform DingTalk. In August 2023, Alibaba Cloud released the weights of Qwen-7B, a 7-billion-parameter bilingual (Chinese–English) model, on platforms such as Hugging Face and ModelScope under a license permitting free commercial use for organizations with fewer than 100 million monthly active users. This was followed by Qwen-14B in September and the larger Qwen-72B in November 2023, along with smaller models (1.8B, 0.5B) and early multimodal models such as Qwen-VL (vision–language) and Qwen-Audio. These releases earned Qwen an early reputation in the open-source community for strong performance relative to model size.

2024: Iterative scaling. In February 2024, Alibaba released the Qwen1.5 series, which extended coverage to a broader range of sizes and added a mixture-of-experts (MoE) variant. The major Qwen2 generation followed in June 2024, with models from 0.5B to 72B parameters trained on substantially larger multilingual corpora, accompanied by a detailed technical report. Through the remainder of 2024, Alibaba expanded the family with Qwen2-Math, Qwen2-VL, the Qwen2.5 series (September 2024, pretrained on roughly 18 trillion tokens), and Qwen2.5-Coder. In November 2024 it introduced QwQ-32B-Preview, an early open reasoning model designed to "think" step by step, and in December 2024 the visual-reasoning model QVQ. The flagship proprietary model Qwen2.5-Max, described as a large mixture-of-experts system trained on more than 20 trillion tokens, was released in January 2025.

2025: Reasoning, agents, and trillion-parameter models. The Qwen3 generation, released in April 2025, introduced hybrid "thinking" and "non-thinking" modes within single models, with open dense models from 0.6B to 32B parameters and MoE models including Qwen3-235B-A22B. Subsequent releases included Qwen3-Coder (July 2025), a very large MoE model specialized for agentic programming, multimodal updates such as Qwen3-VL and Qwen3-Omni, and the proprietary flagship Qwen3-Max, which Alibaba described as exceeding one trillion parameters. In late 2025, Alibaba also launched a dedicated Qwen mobile application for international users, moving beyond its web-based chat interface.

Model architecture and technical characteristics

Qwen models are based on the decoder-only Transformer architecture that underlies most contemporary LLMs. Successive generations incorporated standard modern refinements, including rotary position embeddings (RoPE) for positional encoding, SwiGLU activation functions, root-mean-square layer normalization (RMSNorm), and grouped query attention (GQA) to reduce memory and computation at inference. Context windows lengthened over time, from several thousand tokens in the earliest models to 128K tokens and beyond in later generations, with techniques such as YaRN extending effective context to the million-token range in some variants.

Several characteristics define the family:

• Multilinguality. While bilingual in Chinese and English at inception, later generations were trained on data spanning dozens to more than a hundred languages, making Qwen notably strong in non-English languages relative to many Western peers.
• Scale diversity. The family deliberately spans sizes from roughly 0.5B to hundreds of billions of parameters, enabling deployment on consumer devices as well as data centers.
• Mixture of experts. MoE variants, beginning with Qwen1.5-MoE and continuing through Qwen2, Qwen3, and flagship systems such as Qwen3-Coder, activate only a fraction of total parameters per token, improving efficiency.
• Post-training and reasoning. Models undergo supervised fine-tuning and reinforcement learning from human feedback for instruction following; reasoning-focused releases (QwQ, QVQ, Qwen3 thinking modes) apply reinforcement learning to elicit extended chains of thought.

Model family

The Qwen ecosystem comprises several interrelated lines:

• Base and instruct LLMs across scales (e.g., Qwen, Qwen1.5, Qwen2, Qwen2.5, Qwen3), covering pretraining checkpoints and instruction-tuned chat models.
• Proprietary flagships (Qwen-Max, Qwen2.5-Max, Qwen3-Max), available only through Alibaba's services.
• Reasoning models: QwQ (text) and QVQ (visual reasoning).
• Code models: CodeQwen1.5, Qwen2.5-Coder, and Qwen3-Coder, supporting dozens of programming languages and agentic workflows.
• Mathematics models: Qwen2-Math and Qwen2.5-Math.
• Multimodal models: Qwen-VL/Qwen2-VL/Qwen2.5-VL/Qwen3-VL for image and video understanding, Qwen-Audio and its successors, and the "Omni" line (Qwen2.5-Omni, Qwen3-Omni) unifying text, image, audio, and video input with speech output.

Consumer and developer access is provided through Qwen Chat (chat.qwen.ai), the Tongyi/Qwen apps, and Alibaba Cloud's model platforms (DashScope, the Bailian platform, and the international Model Studio service).

Open-source ecosystem

Open-weight releases have been central to Qwen's strategy. Many models were published under the Apache 2.0 license, among the most permissive licenses available, allowing unrestricted commercial use. Qwen models are distributed on Hugging Face, ModelScope, and GitHub, with ready-made quantizations enabling execution on laptops and smartphones. By 2025, Alibaba reported that the family had accumulated hundreds of millions of downloads and that more than 100,000 derivative models had been created by third parties, figures that placed Qwen among—if not at—the top of the open-source LLM ecosystem by adoption.

The family also served as a base for prominent third-party work. A notable example is the DeepSeek-R1-Distill series released in January 2025, in which DeepSeek fine-tuned Qwen models (and others) on reasoning data distilled from its R1 model, producing small reasoning systems that saw wide uptake. A large fraction of entries on community leaderboards for fine-tuned and specialized models have likewise been built on Qwen weights.

Performance and reception

According to Alibaba's published evaluations, successive Qwen flagships have been competitive with leading frontier models. Qwen2.5-72B performed comparably to other leading open models of similar scale; Qwen2.5-Max was reported to match or exceed contemporaneous systems such as DeepSeek-V3, GPT-4o, and Claude 3.5 Sonnet on selected benchmarks; and Qwen3-235B-A22B achieved results near the frontier in reasoning, mathematics, and coding evaluations. Qwen3-Coder reported strong results on agentic software-engineering benchmarks such as SWE-bench Verified. Independent evaluations generally corroborate that the Qwen family leads the open-model field, particularly in multilingual, mathematical, and coding tasks, though frontier proprietary models have often retained an edge in some agentic and long-horizon tasks.

Commentary has also noted aspects common to models developed in China: Qwen-based chat services have been observed to decline or give constrained answers to politically sensitive prompts, reflecting regulatory requirements applicable to AI services in China. Some early Qwen weights were also modified by third parties into "uncensored" variants, a phenomenon common to open models generally.

Impact and significance

Qwen has had a substantial influence on the global AI landscape. For Alibaba, it anchors a strategy that couples AI research with cloud services, helping Alibaba Cloud compete in the enterprise AI market and, according to press reports in 2025, forming the basis of a partnership with Apple to provide AI features for iPhones sold in China. For the open-source community, Qwen's permissive licensing and broad size range made it a default foundation for research, fine-tuning, and on-device applications worldwide, particularly in Asian, Middle Eastern, and European languages underserved by English-centric models. Within China, Qwen is regarded as one of the most influential model families alongside those of DeepSeek, ByteDance, Zhipu, and Moonshot, and its open releases have been cited as a factor in the rapid diffusion of large-model capability across the domestic industry and the Global South.

As of the mid-2020s, Qwen remains under active development, with Alibaba committing large multi-year investments in AI infrastructure and research, and the family continuing to expand in scale, modality coverage, and agentic capabilities.