CtrlK
Tongyi Qianwen (通义千问, literally "Understanding a Thousand Questions") is an artificial intelligence chatbot and large language model platform developed by Alibaba Cloud, the cloud computing division of Alibaba Group. Launched in April 2023, it serves as Alibaba's flagship AI product, built on top of the Qwen family of large language models. Tongyi Qianwen supports text-based conversation, image understanding, document analysis, code generation, and multimodal interactions. The platform is part of a broader suite of AI tools under the Tongyi brand, including Tongyi Wanxiang for image generation and Tongyi Tingwu for audio and meeting summarization.
Since its initial release, the underlying Qwen models have evolved through several generations, from Qwen 1.0 to Qwen 3.5 as of early 2026. Alibaba has pursued an aggressive open-source strategy, releasing many Qwen model variants under the Apache 2.0 license on Hugging Face and GitHub. In February 2026, Alibaba unified all its AI brands under the single "Qwen" name, retiring the "Tongyi Qianwen" branding for the consumer-facing application.
Alibaba Cloud unveiled Tongyi Qianwen on April 11, 2023, during a press event in Beijing. The announcement positioned it as a direct competitor to ChatGPT and other conversational AI systems that had emerged in late 2022 and early 2023. At launch, the model supported both Chinese and English language capabilities. Alibaba Cloud's enterprise customers in China were given access to Tongyi Qianwen for beta testing, with an initial deployment on DingTalk (Alibaba's workplace messaging platform) and Tmall Genie (Alibaba's smart home assistant).
Alibaba Group CEO Daniel Zhang stated that all of Alibaba's business units would eventually integrate Tongyi Qianwen, signaling the company's intent to weave generative AI throughout its entire ecosystem.
After receiving regulatory approval from China's Cyberspace Administration, Alibaba opened Tongyi Qianwen for public use in September 2023. Users in China could access the chatbot through its official website and a dedicated mobile application. Enterprise customers gained access through the Tongyi API.
At the annual Apsara Conference on October 31, 2023, Alibaba Cloud announced Tongyi Qianwen 2.0. The upgraded version featured an expanded model size and improved alignment techniques. Alibaba highlighted improvements in understanding complex instructions, copywriting, logical reasoning, memorization, and hallucination prevention. Alongside the model update, Alibaba Cloud also introduced industry-specific models tailored for sectors such as finance, legal, and healthcare.
On February 5, 2026, Alibaba Group announced a major brand integration plan, unifying all AI technologies and core brands under the single name "Qwen." The consumer-facing chatbot application was renamed from "Tongyi Qianwen" to "Qwen APP," and the model series was rebranded as "Qwen Large Model." The Chinese brand name became "千问大模型" while the English brand became simply "Qwen." Tongyi Lab continues to exist as the organizational name of Alibaba Group's AI research and development institution. The rebranding aimed to reduce market confusion that had resulted from maintaining multiple brand names across different products.
The primary interface of Tongyi Qianwen is a conversational chatbot that responds to user queries in natural language. It handles tasks such as writing assistance, brainstorming, question answering, translation, and general knowledge queries. The system supports both Chinese and English, along with dozens of additional languages as the underlying Qwen models expanded their multilingual coverage. Qwen3 extended language support to 119 languages and dialects, up from 29 in Qwen2.5.
Through integration with Qwen's vision-language models (such as Qwen-VL and later Qwen2.5-VL), Tongyi Qianwen can analyze and describe uploaded images, answer questions about visual content, and extract text from screenshots or photographs. The Qwen2.5-VL series, released in January 2025, came in variants of 3B, 7B, 32B, and 72B parameters and demonstrated strong performance on visual reasoning benchmarks.
Tongyi Qianwen supports long-document parsing across multiple file formats, including PDF, Word, Excel, and PowerPoint files. The platform can automatically extract key points, generate summaries, create tables from unstructured text, and locate specific clauses within lengthy contracts or reports. Testing showed the system handling documents exceeding 100 pages while producing key point summaries and table generation within minutes.
Alibaba developed specialized coding models within the Qwen family. The Qwen2.5-Coder series, released in late 2024, offered models ranging from 0.5B to 32B parameters. The Qwen2.5-Coder-32B-Instruct model achieved state-of-the-art performance among open-source code models on benchmarks such as EvalPlus, LiveCodeBench, and BigCodeBench, with code generation capabilities comparable to GPT-4o. It supported more than 40 programming languages.
In July 2025, Alibaba released the Qwen3-Coder series, with the flagship Qwen3-Coder-480B-A35B-Instruct model built on a Mixture of Experts (MoE) architecture containing 480 billion total parameters with 35 billion activated per token. This model was designed for agentic AI coding tasks, including generating new code, managing complex workflows, and debugging across entire codebases.
Alibaba Cloud unveiled Tongyi Wanxiang (通义万相, meaning "ten thousand images") at the World Artificial Intelligence Conference in July 2023. The model generates images from text prompts in Chinese and English, supporting styles including watercolor, oil painting, Chinese ink painting, animation, sketch, flat illustration, and 3D cartoon.
Tongyi Wanxiang was built using Composer, Alibaba Cloud's proprietary model architecture that provides control over spatial layout, color palette, and composition while maintaining image quality. The model also supports image-to-image transformation and style transfer, where the content of an original image is preserved while applying the visual style of another picture.
In late 2025, Alibaba Cloud introduced Wanx 2.1, the latest iteration, which expanded capabilities to include high-quality video generation from text input alongside image generation. In January 2026, Alibaba open-sourced Qwen-Image-2512 as a competitive alternative to other image generation AI systems.
| Feature | Description |
|---|---|
| Text-to-Image | Generates images from Chinese and English text prompts |
| Style Transfer | Applies the visual style of one image to the content of another |
| Image-to-Image | Transforms an existing image into a new one with a similar style |
| Video Generation | Wanx 2.1 generates video from text prompts |
| Supported Styles | Watercolor, oil painting, Chinese ink, animation, sketch, 3D cartoon |
Tongyi Tingwu (通义听悟, meaning "listen and understand") is Alibaba Cloud's AI-powered audio and video processing tool. Originally an internal tool called Tingwu used for transcribing project meetings at Alibaba's investment and human resources departments, it was enhanced with large language model capabilities and launched publicly in June 2023.
Tingwu provides real-time transcription of meetings and audio content, automatic speaker identification, chapter division of audio and video content, full-text and per-speaker summaries, key point highlighting, to-do item extraction, and real-time bilingual subtitle translation. The platform integrates with DingTalk, allowing users to automatically transcribe and summarize meetings conducted through Alibaba's workplace communication app.
| Capability | Details |
|---|---|
| Transcription | Real-time audio-to-text with speaker identification |
| Summarization | Full-text summaries, per-speaker viewpoints, key point extraction |
| Chapter Division | Automatic segmentation of audio/video into chapters |
| Translation | Real-time bilingual subtitle translation |
| Action Items | Automatic extraction of to-do items and issues |
| Integration | DingTalk, standalone web and mobile apps |
Alibaba Cloud Model Studio, known in Chinese as Bailian (百炼, meaning "refined through a hundred processes"), is a one-stop platform for building applications using large language models. The platform provides access to over 200 models, including the Qwen language model family, Qwen2-VL for visual understanding, and CosyVoice for speech synthesis.
Bailian offers tools for model fine-tuning, prompt engineering, retrieval-augmented generation (RAG), and agent development. At the 2025 Yunqi Conference, Alibaba Cloud upgraded the Bailian platform to support agent development kits, governance features, and AI guardrails for compliance.
The platform introduced real-time audio and video interaction capabilities in 2024, enabling businesses to build multimodal AI applications without writing code. By 2025, the average daily model call volume on Bailian had increased 15-fold compared to the previous year.
In the first half of 2025, Alibaba Cloud accounted for 35.8% of China's AI cloud market, ranking first among all providers. The Bailian platform serves as a key driver of this market position, providing enterprise customers with the infrastructure to deploy Qwen models for their specific use cases.
Alibaba Group integrated Tongyi Qianwen across multiple business units, making it one of the most broadly deployed enterprise AI systems in China.
DingTalk, Alibaba's workplace communication platform with hundreds of millions of users, was the first product to integrate Tongyi Qianwen. AI-powered features include meeting note summarization, converting spoken conversations to text, email drafting, and business proposal generation. The integration allows users to interact with the AI assistant directly within their existing workflow without switching applications.
Alibaba integrated Tongyi Qianwen into Taobao, its flagship e-commerce platform, as a generative-AI-powered personal assistant. The integration refines product search for users by understanding natural language queries, providing tailored product recommendations, and offering conversational support during the shopping experience. Merchants on Taobao can also use AI tools for product description writing and customer service automation.
Cainiao, Alibaba's global logistics arm, has applied AI capabilities to optimize delivery routing, warehouse operations, and package tracking. The integration of large language model technology assists in customer service automation and logistics planning across Cainiao's international shipping network.
Tongyi Qianwen has also been deployed in Tmall Genie (smart home devices), Amap (navigation and mapping), Youku (video streaming), and Fliggy (travel services). Alibaba Group's leadership identified "User First" and "AI-driven" as the two core strategies for the company, with Tongyi Qianwen serving as the central element of this transformation.
The Qwen (通义千问) model family forms the technical foundation of all Tongyi products. Developed by Alibaba's Tongyi Lab, the models have gone through several major generations.
| Version | Release Date | Key Details |
|---|---|---|
| Qwen (1.0) | August 2023 | Initial 7B parameter open-weight release; 72B and 1.8B released December 2023 |
| Qwen1.5 | February 2024 | Models from 0.5B to 110B parameters |
| Qwen2 | June 2024 | Major upgrade; released with open weights for smaller models, proprietary for largest |
| Qwen2.5 | September 2024 | Added 3B, 14B, and 32B sizes; trained on 18 trillion tokens |
| Qwen2.5-Max | January 2025 | MoE architecture; trained on over 20 trillion tokens; proprietary |
| Qwen2.5-VL | January 2025 | Vision-language models in 3B, 7B, 32B, and 72B variants |
| Qwen2.5-Omni | March 2025 | End-to-end multimodal model handling text, images, audio, and video |
| QwQ-32B | March 2025 | 32.5B parameter reasoning model; 131K context window |
| Qwen3 | April 2025 | Dense (0.6B to 32B) and MoE (30B-A3B, 235B-A22B) models; 119 languages; Apache 2.0 |
| Qwen3-Coder | July 2025 | 480B-A35B flagship coding model |
| Qwen3-Max | September 2025 | Proprietary top-tier model |
| Qwen3-Next | September 2025 | Open-weight release under Apache 2.0 |
| Qwen3-Omni | September 2025 | Open-source multimodal model generating text, images, audio, and video |
| Qwen3-Max-Thinking | January 2026 | Reasoning-enhanced proprietary model |
| Qwen3.5 | February 2026 | 397B-A17B MoE flagship; native vision-language; 201 languages; open-weights |
Released on April 28, 2025, the Qwen3 family includes both dense models (0.6B, 1.7B, 4B, 8B, 14B, and 32B parameters) and Mixture-of-Experts models (Qwen3-30B-A3B and the flagship Qwen3-235B-A22B). The models were trained on 36 trillion tokens in 119 languages and dialects. A notable innovation in Qwen3 is the unified thinking/non-thinking mode, which integrates complex multi-step reasoning and rapid context-driven responses into a single model, eliminating the need to switch between separate chat and reasoning models.
Qwen3 showed density improvements of roughly 50%, meaning that Qwen3-1.7B, Qwen3-4B, Qwen3-8B, Qwen3-14B, and Qwen3-32B matched the performance of Qwen2.5-3B, Qwen2.5-7B, Qwen2.5-14B, Qwen2.5-32B, and Qwen2.5-72B, respectively. All Qwen3 models were released under the Apache 2.0 license.
Released on February 16, 2026, Qwen3.5 is a native vision-language model series built on a Sparse MoE architecture. The flagship Qwen3.5-397B-A17B packs 397 billion total parameters but activates only 17 billion per token. The series includes eight models ranging from 0.8B to 397B parameters, with additional sizes released on February 24, 2026, including Qwen3.5-122B-A10B (supporting 1M+ context lengths), Qwen3.5-35B-A3B, and Qwen3.5-27B. Qwen3.5 expanded multilingual support to 201 languages and dialects.
Alibaba introduced QwQ (Qwen-with-Questions) in November 2024 as a reasoning model designed to compete with OpenAI's o1 series. The initial QwQ-32B-Preview was released under the Apache 2.0 license. In March 2025, the improved QwQ-32B was open-sourced, featuring 32.5 billion parameters and a 131,072-token context window. The model was trained using reinforcement learning and structured self-questioning, showing performance comparable to DeepSeek-R1 on math and coding benchmarks. Alibaba's stock price rose more than 8% following the QwQ-32B release.
Released in March 2025, Qwen2.5-Omni is an end-to-end multimodal model that perceives text, images, audio, and video while generating both text and natural speech responses in a streaming manner. The model uses a novel Thinker-Talker architecture: the Thinker component processes and understands inputs from all modalities, while the Talker component generates speech output. It also employs TMRoPE (Time-aligned Multimodal RoPE) to synchronize video and audio timestamps. Variants include the 7B and 3B parameter sizes, with 4-bit quantized versions available that reduce GPU memory consumption by over 50%.
Alibaba has been one of the most active contributors to the open-source AI ecosystem globally. The company has released more than 100 open-weight models, accumulating over 40 million downloads across Hugging Face, GitHub, and ModelScope.
Most Qwen model variants are distributed under the Apache 2.0 license, which permits commercial use, modification, and redistribution. However, Alibaba follows a tiered strategy: the most advanced proprietary models (such as Qwen2.5-Max and Qwen3-Max) remain closed-source and are served through Alibaba Cloud, while smaller and mid-sized models are open-sourced to encourage community adoption and ecosystem development.
Over 130,000 derivative models based on Qwen have been developed by the global open-source community on Hugging Face. Qwen models have repeatedly topped the Hugging Face Open LLM Leaderboard. In April 2025, Alibaba's Qwen2.5-Omni topped Hugging Face's open-source AI model trending list.
The Tongyi Qianwen team from Alibaba won the Best Paper Award at NeurIPS 2025 for their paper "Attention Gating Makes Better Foundation Models," the only Chinese entry among the four winning papers at the conference. In November 2025, Singapore's National AI Programme adopted the open-source Qwen architecture for its Southeast Asian language large model project.
Alibaba maintains several public repositories for the Qwen model family:
| Repository | Description |
|---|---|
| QwenLM/Qwen | Original Qwen (1.0) models and chat variants |
| QwenLM/Qwen2.5 | Qwen2.5 series models |
| QwenLM/Qwen3 | Qwen3 dense and MoE models |
| QwenLM/Qwen3.5 | Qwen3.5 native vision-language models |
| QwenLM/Qwen3-Coder | Specialized code generation models |
| QwenLM/Qwen3-VL | Multimodal vision-language models |
| QwenLM/QwQ | Reasoning model series |
In February 2025, Alibaba Chairman Joe Tsai confirmed at the World Governments Summit in Dubai that Alibaba would partner with Apple to provide AI features for iPhone users in China. Apple had been unable to roll out its Apple Intelligence features in China due to the country's stringent regulations requiring large language models to receive government approval before commercial deployment.
Apple evaluated models from several Chinese companies, including ByteDance, Tencent, and DeepSeek, before selecting Alibaba's Qwen. The partnership was driven by Qwen's advanced capabilities, its proven integration potential, and DeepSeek's relative lack of the manpower and experience needed to support a large customer like Apple. The deal was considered strategically significant for both companies: Apple needed a local AI partner to boost iPhone sales in China (the world's largest smartphone market), while Alibaba gained validation of its AI capabilities from one of the world's most valuable technology companies.
Tongyi Qianwen operates in one of the most competitive AI markets in the world. China's AI chatbot landscape includes major players from nearly every large technology company in the country.
| Competitor | Company | Notable Details |
|---|---|---|
| Doubao | ByteDance | Leading in consumer market with integration into Douyin; aggressive pricing |
| ERNIE Bot | Baidu | Early mover in Chinese AI chatbots; ERNIE 4.5 Turbo released in 2025 |
| Hunyuan | Tencent | Integrated into WeChat and Tencent Cloud services |
| DeepSeek | High-Flyer | Open-source reasoning models; DeepSeek-R1 and V3 gained global attention |
| ChatGLM | Zhipu AI | GLM series models; backed by Tsinghua University |
| Kimi | Moonshot AI | Known for long context window capabilities |
| SparkDesk | iFlytek | Strong in speech recognition and education |
The Chinese AI chatbot market has seen rapid shifts in user adoption. ByteDance's Doubao emerged as a consumer market leader in late 2024, benefiting from integration with Douyin (the Chinese version of TikTok) and an aggressive pricing strategy. However, the competitive landscape remained fluid. Following the Qwen brand unification in early 2026, Alibaba's Qwen APP saw daily active users reach 73.52 million during the Spring Festival holiday, processing nearly 200 million "one-sentence order" instructions.
Baidu's ERNIE Bot, which had been an early mover in the market, saw declining downloads through 2024 as competition intensified. In response, Baidu made ERNIE Bot freely accessible to all users in early 2025.
Alibaba differentiates itself through its dual strategy of maintaining competitive proprietary models while simultaneously leading in open-source model releases, its deep integration across the Alibaba business ecosystem, and the Bailian platform that serves enterprise customers.
The Qwen models are developed by Tongyi Lab, Alibaba Group's central AI research institution. Tongyi Lab was established in 2023 by consolidating several AI teams from DAMO Academy, Alibaba's research division.
Zhou Jingren serves as CTO of Alibaba Cloud and head of Tongyi Lab. He holds a Ph.D. in Computer Science from Columbia University, previously worked at Microsoft Research, and joined Alibaba in 2015. He served as Chief Scientist at Alibaba Cloud and Deputy Director of DAMO Academy before becoming CTO in 2022.
In early March 2026, Lin Junyang, the technical lead of the Qwen large-model team, stepped down from his role. Following his departure, Zhou Jingren assumed direct oversight of the Qwen model development. The organizational restructuring involved splitting the Qwen team from a vertically integrated structure into a horizontally organized system with separate teams for pre-training, post-training, text, multimodal, and other functions.
The pace of model releases accelerated through 2025 and into 2026. Key milestones include:
Alibaba announced its first AI glasses product under the "Qwen" brand, debuting at the 2026 Mobile World Congress (MWC) with pre-orders starting March 2, 2026. This marked Alibaba's entry into consumer AI hardware beyond software and cloud services.
Alibaba Cloud reported that the daily model call volume on the Bailian platform increased 15-fold over the course of 2024-2025. In the first half of 2025, Alibaba Cloud held 35.8% of China's AI cloud market, maintaining its leading position.
| Specification | Details |
|---|---|
| Developer | Alibaba Cloud / Tongyi Lab |
| Initial Release | April 11, 2023 |
| Latest Model Generation | Qwen3.5 (February 2026) |
| Largest Open Model | Qwen3.5-397B-A17B (397B total, 17B active) |
| Largest Coding Model | Qwen3-Coder-480B-A35B (480B total, 35B active) |
| Training Data (Qwen3) | 36 trillion tokens |
| Languages Supported (Qwen3.5) | 201 languages and dialects |
| License (Open Models) | Apache 2.0 |
| Platform | Alibaba Cloud Model Studio (Bailian) |
| Consumer App | Qwen APP (formerly Tongyi Qianwen) |
| Access | chat.qwen.ai, mobile apps, API |