Qwen

Overview

Tongyi Qianwen (Qwen) is a series of large language models and AI assistants launched by Alibaba, leading the comprehensive capabilities of domestic large models. It performs well in Chinese understanding, long text processing, logical reasoning, and coding abilities, and the consumer app is completely free.

Latest flagship: Qwen3.8-Max (2.4T parameters), which moved from 40 to 45 on the Artificial Analysis leaderboard, a gain Alibaba attributes to self-improvement running 33 effective iterations with no human involvement (2026/9/22). The Qwen3.8-Max-0902 release (2026/9/2) debuted at the top of the overall Code Arena: WebDev leaderboard with a score of 1,691, and became the highest-scoring model on the Pareto frontier at a blended price of $5/MToken, available for trial on QwenCloud. The Qwen3.8-Flash open-source multimodal MoE model (2026/8/26) serves as an early preview of the Qwen4 architecture with open weights, 125B total parameters activating 6B per token, training costs just 1/9 of Qwen3.7-Plus while outperforming it comprehensively, priced at $0.16/1M input tokens and $0.47/1M output tokens, with a native context of 262K expandable to 1M. The Qwen3.8 series was open-sourced (2026/8/14), including the native multimodal dense Qwen3.8-27B, which surpasses Qwen3.7-Plus, natively supports a 262K context extendable to 1M via YaRN under Apache 2.0, alongside the open weights of the Max-level Qwen3.8-2.4T-A95B. At the 2026 Yunqi conference (2026/9/22), Alibaba shared that Qwen4 has entered training on a next-generation architecture, with Qwen4.5 and Qwen5 planned to scale from 5 to 10 trillion parameters. The Qwen series has an active open-source ecosystem with tens of millions of global downloads. Additionally, the latest information on the official website shows that Tongyi Qianwen has launched Qwen3-VL-Flash and other vision-language models, as well as the Wan2.6 series models (such as Wan2.6-R2V, Wan2.6-I2V, Wan2.6-T2V, Wan2.6-T2I, Wan2.6-Image) and the Wanxiang visual generation large model, which possess multimodal generation capabilities for images, videos, and sound, achieving leading levels in image quality, semantic understanding, motion amplitude, adherence to physical laws, and artistic texture. Wan3.0 took first place on both the Artificial Analysis text-to-video and video-editing leaderboards (2026/9/22).

Key Features

• Ultra-long Context: Supports a context window of up to 1 million tokens, capable of processing an entire book at once
• Multimodal Understanding: Supports understanding and analysis of text, images, audio, video, and documents; newly added vision-language model Qwen3-VL-Flash and Wanxiang visual generation large model, supporting image, video, and sound generation
• Web Search: Real-time online search for the latest information, with answers including source links
• Document Analysis: Upload PDF, Word, and other documents for intelligent analysis and Q&A
• Coding Ability: Supports code generation, explanation, and debugging in multiple programming languages, leading in multiple programming benchmarks; Qwen3.8-Max-0902 debuted at the top of the overall Code Arena: WebDev leaderboard with a score of 1,691 (2026/9/2)
• Open-source Model Ecosystem: The Qwen series models are fully open-source, with tens of millions of global downloads and top community activity; the Qwen3.8 series was open-sourced on 2026/8/14, including Qwen3.8-27B (native multimodal dense, 262K context extendable to 1M via YaRN, Apache 2.0) and the Max-level Qwen3.8-2.4T-A95B weights
• Hybrid Attention Mechanism (Plus): New architecture brings a 19x increase in inference throughput and a 60% reduction in deployment costs
• Offline Version (Qwen3-Max Offline Version): Built-in on-device inference version in the Qianwen App, fully local operation without internet, suitable for basic Q&A and text continuation
• Qianwen App Super Ecosystem: Fully integrated with Alibaba's ecosystem including Taobao, Alipay, Fliggy, and Amap, enabling AI to directly 'handle tasks'; a major Qwen App update (2026/8/7) added deep research, scheduled tasks, an office assistant (with paid capacity expansion), an agent marketplace and voice calls, supporting the flagship Qwen3.8-MAX with 2.4T parameters
• Wan2.6 Series Models: Includes Wan2.6-R2V (video character reference generation), Wan2.6-I2V (intelligent multi-shot narrative), Wan2.6-T2V (natural audio-visual synchronization), Wan2.6-T2I (strong instruction-following ability), Wan2.6-Image (mixed text-image output), etc., enhancing multimodal generation capabilities
• Wanxiang Visual Generation Large Model: Trained through a native multimodal unified framework, possessing multimodal generation capabilities for images, video, and sound, achieving leading levels in image quality, semantic understanding, motion amplitude, adherence to physical laws, and artistic texture
• Qwen3.8-Flash Open-Source MoE: Released 2026/8/26 as an early preview of the Qwen4 architecture with open weights; 125B total parameters activating 6B per token, training costs just 1/9 of Qwen3.7-Plus while outperforming it comprehensively; production API priced at $0.16/1M input tokens and $0.47/1M output tokens, native context 262K expandable to 1M
• Qwen-Image-3.0-Pro: Launched on Qwen Cloud (2026/8/5), ranking first among Chinese models and second among mainstream models on the Arena text-to-image leaderboard; supports 4.5k-token prompts, 10px-level text rendering, and 12 languages; Pro starts at $0.04/image and Standard at $0.03/image
• Qwen-Audio-3.1 Family: Released 2026/9/23, upgrading the ASR, TTS and Realtime lines and adding TTS-Next for audio creation and ASR-Next for audio understanding, five models covering comprehension, generation, interaction and creation; prices fell across the line, roughly 70% for TTS, about 85% for Realtime and up to 95% for ASR; Realtime listens while speaking and can be interrupted at any point, slowing its pace and responding with empathy when it detects a low mood

Use Cases

• Chinese writing, translation, and content creation
• Reading comprehension and summarization of long documents
• Programming assistance and technical Q&A
• Daily information queries and knowledge Q&A
• Base model for enterprise private deployment (open-source version)
• Scenarios requiring offline AI operation (on-device version)
• Deep integration with Alibaba's ecosystem for e-commerce/travel/payment scenarios
• Visual content generation: image, video, and sound creation (Wanxiang visual generation large model)
• Video character reference generation, intelligent multi-shot narrative (Wan2.6 series)
• Audio understanding, generation, interaction and creation (Qwen-Audio-3.1 family)

Pros

• Consumer

Summary

Tongyi Qianwen (Qwen) is suitable for users who require high-quality Chinese comprehension, long-text processing, multimodal generation, and deep integration with the Alibaba ecosystem, such as content creators, developers, enterprises, and e-commerce/travel scenarios. Its core advantages include the flagship Qwen3.8-Max with 2.4T parameters, the Qwen3.8-Max-0902 topping Code Arena: WebDev and leading the Pareto frontier at $5/MToken, the open-source Qwen3.8 series including Qwen3.8-27B and Qwen3.8-2.4T-A95B, the Qwen3.8-Flash multimodal MoE preview of the Qwen4 architecture, support for ultra-long context (1 million tokens), multimodal understanding and generation (including the Wan2.6 series, Wanxiang visual model, and Qwen-Image-3.0-Pro), an active open-source ecosystem, and a completely free consumer-side app. Additionally, it offers offline versions and efficient deployment enabled by a hybrid attention mechanism.

Version History

• Qwen releases the Qwen-Audio-3.1 family, upgrading ASR, TTS and Realtime (2026-09-23): Qwen-Audio-3.1 upgrades the ASR, TTS and Realtime lines and adds two new models, TTS-Next for audio creation and ASR-Next for audio understanding, covering comprehension, generation, interaction and creation with five models in total. Prices drop across the board, roughly 70% off for TTS and 85% off for Realtime.
• Ships the five-model Qwen-Audio-3.1 family with across-the-board price cuts (2026-09-23): Qwen released Qwen-Audio-3.1, upgrading ASR, TTS and Realtime while adding the audio creation model TTS-Next and the audio understanding model ASR-Next, five models covering understanding, generation, interaction and creation. Prices fell across the line, roughly 70% for TTS, about 85% for Realtime and up to 95% for ASR. Realtime listens while speaking and can be interrupted at any point, slowing its pace and responding with empathy when it detects a low mood.
• Yunqi 2026: Qwen4 enters training, Qwen-Image-3.1 and Qwen-Audio-3.1 progress (2026-09-22): Alibaba shared its model roadmap at the 2026 Yunqi conference: Qwen4 has entered training on a next-generation architecture, with Qwen4.5 and Qwen5 planned to scale from 5 to 10 trillion parameters. Qwen-Image-3.1 is expected to approach the GPT-Image series, Qwen-Audio-3.1 ranks first in China across ASR, TTS and realtime voice, and the world model HappyOyster-2.0-Preview made its debut. Wan3.0 took first place on both the Artificial Analysis text-to-video and video-editing leaderboards, and the next video model is set for November. Qwen3.8-Max moved from 40 to 45 on the Artificial Analysis leaderboard, which Alibaba says came from self-improvement running 33 effective iterations with no human involvement.
• Qwen3.8-Max-0902 tops Code Arena and leads the Pareto frontier at $5/MTok (2026-09-02): Tongyi Qianwen released Qwen3.8-Max-0902, debuting at the top of the overall leaderboard in Code Arena: WebDev with a score of 1,691, and becoming the highest-scoring model on the Pareto frontier at a blended price of $5/MToken. It is now available for trial on QwenCloud.
• Qwen3.8-Flash goes open source, with a preview of the Qwen4 architecture (2026-08-26): Tongyi Qianwen released Qwen3.8-Flash, a multimodal MoE model, serving as an early preview of the Qwen4 architecture with open weights. The model has a total of 125B parameters, activating only 6B per token, with training costs just 1/9 of Qwen3.7-Plus, while outperforming it comprehensively. The production API is priced at $0.16/1M input tokens and $0.47/1M output tokens, with a native context of 262K, expandable to 1M.
• Qwen open-sources the Qwen3.8 model series (2026-08-14): Tongyi Qianwen has fulfilled its promise by open-sourcing the Qwen3.8 series models. Among them, Qwen3.8-27B is a native multimodal dense model, which, with only 27B parameters, comprehensively surpasses Qwen3.7-Plus. It natively supports a 262K context, can be extended to 1M tokens via YaRN, and is licensed under Apache 2.0. The open weights of the Max-level Qwen3.8-2.4T-A95B have also been released simultaneously.
• Major Qwen App update plus Qwen3.8-MAX (2026-08-07): Deep research, scheduled tasks, an office assistant (with paid capacity expansion), an agent marketplace and voice calls are now live, supporting the flagship Qwen3.8-MAX with 2.4T parameters; office features come with paid quotas, and the company plans to take a cut from major open-source commercial users
• Qwen-Image-3.0-Pro launches on Qwen Cloud (2026-08-05): Alibaba's Tongyi Qianwen releases Qwen-Image-3.0-Pro and Standard, now available on Qwen Cloud. The model ranks first among Chinese models and second among mainstream models on the Arena text-to-image leaderboard, supporting 4.5k-token prompts, 10px-level text rendering, and 12 languages. The Pro version starts at $0.04/image, and the Standard version starts at $0.03/image.