Qwen is Alibaba’s family of artificial intelligence models and products. It began as Tongyi Qianwen, a Chinese and English large language model announced in April 2023.
The project soon developed into an open-weight model family for general language work, coding, mathematics, vision, audio, reasoning, image generation, and AI agents.
Qwen-Image-2.1-Turbo arrived on October 9, 2026, with downloadable weights for image generation and editing in eight denoising steps. It uses the same 7-billion-parameter visual generation architecture as Qwen-Image-2.1 and supports 2K output.
The original Qwen-Image-2.1 launched on September 20 with native transparent-image generation and editing and support for up to 10 reference images. Pro and Turbo are also available as hosted APIs. Both downloadable checkpoints use the Qwen Research License, which requires a separate license for commercial use.
Alibaba said Qwen4 is in training and has not announced a release date. It also said Qwen-Image 3.1 is planned for later in 2026. Qwen3.8-Flash-Next remains the latest open-weight general model.
This timeline follows Qwen’s major model releases, product launches, and roadmap milestones from April 2023 through October 2026.
Last Updated: October 9, 2026
Qwen in Brief
| Milestone | Details |
|---|---|
| First public announcement | Tongyi Qianwen was unveiled on April 11, 2023. |
| First open-weight release | Qwen-7B and Qwen-7B-Chat arrived on August 3, 2023. |
| Qwen3 release | Qwen3 was announced on April 28, 2025, followed by public repository releases dated April 29. |
| Latest general generation | Qwen3.8 spans Max, 27B, Flash, Flash-Next, Omni-Flash, Omni-Flash-Realtime, and LiveTranslate across hosted and open-weight use. |
| Latest Qwen release | Qwen-Image-2.1-Turbo, released on October 9, 2026, adds eight-step image generation and editing with downloadable weights. Pro and Turbo image APIs are also available. |
| Latest product platform | Qwen Intelligence, announced on September 22, 2026, gives phone makers access to a Qwen-powered agent platform for complex tasks across mobile apps. |
| Qwen4 status | Alibaba said Qwen4 is in training as of September 22, 2026. It has not announced a public release date. |
| Latest open-weight general model | Qwen3.8-Flash-Next, released on August 26, 2026, is a multimodal MoE model and an architecture preview for Qwen4. |
| Latest Max snapshot | Qwen3.8-Max-0902, released on September 2, 2026, updated the hosted flagship’s coding, multi-tool agent, and visual capabilities. |
| Qwen Code launch | Qwen Code launched with Qwen3-Coder on July 22, 2025. |
Qwen Timeline at a Glance
| Milestone | Date and significance |
|---|---|
| Qwen-Image-2.1-Turbo and hosted image APIs | October 9, 2026 The accelerated open-weight checkpoint generates and edits 2K images in eight denoising steps. Qwen announced Pro and Turbo API availability. |
| Tongyi Qianwen announced | April 11, 2023 Alibaba Cloud publicly introduced the large language model that became the Qwen family. |
| First open-weight Qwen model | August 3, 2023 Qwen-7B and Qwen-7B-Chat began Qwen’s public model release history. |
| Qwen-VL | August 22, 2023 The first vision-language branch added image understanding, text reading, and visual grounding. |
| Qwen1.5 | February 4, 2024 A wider size range and direct Transformers support reduced setup and hardware barriers. |
| Qwen2 | June 7, 2024 The second generation expanded multilingual support, context length, coding, and mathematics. |
| Qwen2.5 | September 19, 2024 General, Coder, and Math models arrived as a coordinated family. |
| QwQ-32B-Preview | November 28, 2024 Qwen opened a separate research line for longer, step-by-step reasoning. |
| Qwen2.5-Omni | March 27, 2025 One end-to-end model could process text, images, audio, and video and return text or speech. |
| Qwen3 | April 28, 2025 Thinking and non-thinking behavior entered the main general-purpose model family. |
| Qwen3-Coder and Qwen Code | July 22, 2025 The agent-oriented coding model launched alongside an open-source terminal coding tool. |
| Qwen-Image | August 4, 2025 Qwen gained a dedicated open-weight image generation and editing model. |
| Qwen3-VL and Qwen3-Omni | September 2025 The visual and full multimodal branches added deeper reasoning, agent use, and real-time speech. |
| Qwen3.5 | February 15, 2026 The main open-weight family became natively multimodal and more focused on agents. |
| Qwen3.6 | April 2026 Hosted and open-weight releases strengthened agentic coding and multimodal work. |
| Qwen3.7 | May-June 2026 Qwen3.7-Max and Qwen3.7-Plus moved the hosted flagship line toward long-running multimodal agents. |
| Qwen3.8-Max-Preview | July 19, 2026 The hosted preview became available through Alibaba’s Token Plan, Qoder, and QoderWork before the full Qwen3.8-Max release. |
| Qwen-Image-3.0 | July 22, 2026 The third-generation image model added complex layouts, precise small-text rendering, 12-language output, and live web retrieval for design and content generation. |
| Qwen-Audio-3.0-TTS | July 23, 2026 Flash and Plus variants added expressive inline controls, natural-language direction, 16-language output, and long-form synthesis up to three minutes. |
| Qwen-Audio-3.0-ASR-Flash | July 31, 2026 The speech-recognition release improved context consistency and domain-term recognition and added custom hotwords and structured transcripts. |
| Qwen3.8-Max | August 2, 2026 The hosted 2.4-trillion-parameter flagship targets long-horizon coding, professional work, and multimodal planning. |
| Qwen3.8-2.4T-A95B | August 8, 2026 Qwen released the Max-class foundation model as a downloadable, text-only MoE checkpoint with 95 billion active parameters. |
| Qwen3.8-27B | August 14, 2026 A dense, native vision-language model brought Qwen3.8 reasoning and agent capabilities to a smaller Apache 2.0 checkpoint. |
| Qwen3.8-Flash-Next | August 26, 2026 The open-weight multimodal MoE introduced Qwen Sparse Attention, Gated Residual, and n-gram embedding as an early Qwen4 architecture preview. |
| Qwen3.8-Max-0902 | September 2, 2026 An upgraded hosted Max snapshot improved engineering-scale coding, multi-tool collaboration, and visual understanding. It retains a one-million-token context window. |
| Qwen4 training and Qwen-Image 3.1 roadmap | September 22, 2026 Alibaba said Qwen4 is in training and Qwen-Image 3.1 is planned for later in 2026. It did not announce a Qwen4 release date. |
| Qwen Intelligence | September 22, 2026 Alibaba introduced a smartphone-focused agent platform for phone makers, designed to handle complex tasks across apps. |
| Qwen-Audio-3.1 and TTS-Next | September 22, 2026 The updated speech suite adds ASR and TTS models. TTS-Next can generate dialogue, sound effects, and ambient audio together. |
| Qwen3.8-Omni-Flash-Realtime | September 21, 2026 The hosted model supports live audio-video interaction with text and audio output, multichannel audio, and remote MCP tools. |
| Qwen-Image-2.1 | September 20, 2026 The open-weight image model unifies generation and editing, with native transparency and up to 10 reference images. |
| Qwen-Audio-3.1-Realtime-Plus | September 20, 2026 The duplex speech model adds eight new system voices, a 262,144-token context window, function calling, web search, and voice cloning. |
| Qwen3.8-Omni-Flash | September 18, 2026 The hosted Omni model combined text, image, audio, and video understanding with thinking, tool calling, web search, and a one-million-token context window. |
| Qwen3.8-LiveTranslate | September 18, 2026 The real-time interpretation model added speaker separation, synchronized source and translated text, long-context disambiguation, and speech output that preserves speaker timbre. |
Qwen Model Release History
| Release | What changed |
|---|---|
| Qwen-Image-2.1-Turbo October 9, 2026 Image generation and editing / open-weight checkpoint | The accelerated Qwen-Image-2.1 checkpoint uses the same 7B visual generation architecture for 2K image generation and natural-language editing in eight denoising steps. Diffusers loads its recommended sampling schedule through QwenImage21Pipeline. The weights are available on Hugging Face and ModelScope under the Qwen Research License; commercial use requires a separate license. Qwen also announced that Pro and Turbo hosted APIs are live. |
| Qwen-Audio-3.1-TTS-Next September 22, 2026 Audio generation / hosted model | The September 22 speech update also includes refreshed ASR and TTS services. TTS-Next accepts text and reference audio and generates speech, dialogue, sound effects, and ambience together. Podcast output can run up to four minutes. Alibaba Cloud Model Studio provides the inference service. |
| Qwen3.8-Omni-Flash-Realtime September 21, 2026 Real-time multimodal conversation / hosted model | The model supports live audio and video interaction with text and audio output. It adds multichannel audio, video aggregation, and remote MCP tools, with access through WebSocket, WebRTC, and AOQ. |
| Qwen-Image-2.1 September 20, 2026 Image generation and editing / open-weight model | A unified image model with a 7B visual generation component. It generates and edits regular or transparent RGBA images, extracts subjects from photos, and accepts up to 10 reference images. Circles, annotations, or masks can specify local edits. Mixed-granularity attention and prefix KV caching reduce computation. The weights use the Qwen Research License, with a separate license required for commercial use. |
| Qwen-Audio-3.1-Realtime-Plus September 20, 2026 Duplex speech conversation / hosted model | The real-time model adds eight new system voices and supports a 262,144-token context window, function calling, web search, and voice cloning. |
| Qwen3.8-Omni-Flash September 18, 2026 Omni / hosted multimodal agent model | Qwen3.8-Omni-Flash accepts text, images, audio, and video and returns text through Chat Completions or Responses. Thinking is enabled by default with adjustable reasoning effort. The model supports function calling, web search, a one-million-token context window, context caching, and two-channel stereo or four-channel FOA spatial-audio input. The release focuses on audio-video understanding and text generation. |
| Qwen3.8-LiveTranslate September 18, 2026 Real-time multilingual interpretation / hosted API | Qwen3.8-LiveTranslate uses an interleaved audio-text architecture for simultaneous interpretation. It separates speakers in real time, keeps source and translated text synchronized, uses conversation history to disambiguate names and terminology, and preserves speaker timbre in translated speech. The API model qwen3.8-livetranslate-flash-realtime accepts audio and image input, recognizes 60 source languages, and supports speech output in 29 languages. Qwen reports average lagging of 2.3 seconds, down from 2.8 seconds in the previous generation. |
| Qwen-Drive-1.0 September 3, 2026 Autonomous driving / vision-language research model | Built on Qwen3.5-4B, Qwen-Drive-1.0 extends a general vision-language model into driving perception and motion planning. It adds a BEV perception head for 3D detection, semantic occupancy, and map segmentation, plus a Planning Expert that generates future vehicle trajectories. |
| Qwen3.8-Max-0902 September 2, 2026 General / hosted flagship snapshot | This upgraded Qwen3.8-Max snapshot improved engineering-scale coding, long-horizon autonomous development, multi-tool agent collaboration, and visual understanding for charts, documents, and multimodal perception. It retains the one-million-token context window, thinking mode, and built-in tool ecosystem. |
| Qwen3.8-Flash-Next August 26, 2026 General / open-weight multimodal architecture preview | Qwen opened the weights of a multimodal MoE model that previews the architecture planned for Qwen4. Its 125-billion-parameter language model activates 6 billion parameters per token and is supplemented by 51 billion n-gram embedding parameters and a 4-billion-parameter multi-token prediction layer. Qwen Sparse Attention selects relevant micro-blocks for long-context work. Gated Residual expands the residual stream into four controlled branches. The model supports text, images, and video, uses a native 262,144-token context window, and can be extended to one million tokens. Qwen3.8-Flash is the production counterpart with a one-million-token context window and built-in tools. At release, Qwen3.8-Flash-Next cost $0.16 per million input tokens and $0.47 per million output tokens. API access was not yet available. |
| Qwen3.8-27B August 14, 2026 General / open-weight multimodal dense model | This 27-billion-parameter dense model brought Qwen3.8’s coding, professional-work, research, and long-horizon agent improvements to a smaller downloadable checkpoint. It natively understands text, images, and video, can switch between thinking and direct responses, and supports a 262,144-token context window that can be extended to one million tokens. The model uses the Apache 2.0 license. |
| Qwen3.8-2.4T-A95B August 8, 2026 General / open-weight thinking model | Qwen released its first downloadable Max-class model in Transformers format. The hybrid Gated DeltaNet and Gated Attention MoE has 2.4 trillion total parameters, 95 billion active parameters, 92 layers, and 512 experts. It accepts text only, always uses thinking mode, and has a native 262,144-token context window that can be extended to 1,010,000 tokens. Its 213 BF16 weight shards total about 4.89 TB in decimal units. The full checkpoint requires data-center-scale storage and serving hardware. Qwen recommends SGLang, vLLM, or TokenSpeed for dedicated serving. |
| Qwen3.8-Max August 2, 2026 General / hosted flagship | Qwen3.8-Max is a 2.4-trillion-parameter model for coding and long-running agent work. It supports autonomous coding for more than 10 days, long-horizon tasks such as 500-plus chip-design turns and a 365-day e-commerce strategy, and multimodal feedback for planning and self-correction. Pricing is $2 per million input tokens, $6 per million output tokens, and $0.25 per million tokens for implicit caching. The service is based on Qwen3.8-2.4T-A95B and adds vision input, non-thinking responses, built-in tools, and one-million-token context by default. |
| Qwen-Audio-3.0-ASR-Flash July 31, 2026 Speech recognition / API | Qwen-Audio-3.0-ASR-Flash adds context consistency, domain-term recognition, custom hotwords, and speech polishing for structured transcripts. Medical-term recall reached 95.36% and industrial-term recall reached 93.24% in internal tests. It is available through streaming, file-transcription, and non-real-time API paths. |
| Qwen-Audio-3.0-TTS July 23, 2026 Speech generation / API | Qwen-Audio-3.0-TTS comes in Flash for real-time interaction and Plus for higher-quality generation. Inline tags such as [whisper], [angry], [breaths], and [laughs] control delivery, and natural-language instructions control the overall speaking style. It supports 16 languages, noisy reference audio, and up to three minutes of long-form speech in one pass. |
| Qwen-Image-3.0 July 22, 2026 Image generation | Qwen-Image-3.0 is the third-generation image model. It accepts prompts up to 4.5K tokens, handles newspapers, storyboards, exam papers, and 3×3 infographics in one pass, renders text down to 10 pixels, supports full LaTeX pages, 12 languages, 100-plus art styles, realistic interfaces, and live web retrieval. |
| Qwen3.8-Max-Preview July 19, 2026 General / hosted preview | The hosted preview became available through Alibaba’s Token Plan, Qoder, and QoderWork. Qwen3.8 is a continuously evolving 2.4-trillion-parameter system. |
| Qwen3.7-Max visual snapshot June 10, 2026 General / multimodal agent | Updated the hosted Max model with visual understanding and multimodal agent interaction. |
| Qwen3.7-Plus June 1, 2026 General / multimodal agent | Hosted vision-language agent model based on the Qwen3.7 generation, with stronger visual understanding and agent work. |
| Qwen3.7-Max May 20, 2026 General / agent | Proprietary flagship for coding, office automation, tool use, and autonomous workflows. |
| Qwen3.6 open models April 2026 General / multimodal | Open-weight MoE and dense models brought Qwen3.6’s coding, reasoning, and visual capabilities to local deployment. |
| Qwen3.6-Max-Preview April 18, 2026 General / agent | Hosted preview with stronger coding agents, knowledge, and instruction following than Qwen3.6-Plus. |
| Qwen3.6-Plus April 2, 2026 General / multimodal agent | Hosted production model with a one-million-token context window and an emphasis on repository work, visual coding, and multi-step execution. |
| Qwen3.5 February 15, 2026 General / native multimodal | Introduced a natively multimodal main generation for text, images, video, reasoning, coding, and agent tasks. The first open model was Qwen3.5-397B-A17B. |
| Qwen3-Coder-Next February 2, 2026 Coder / open-weight agentic coding model | Qwen3-Coder-Next is an open-weight coding model built on Qwen3-Next-80B-A3B-Base for coding agents and local development. Its training emphasizes executable software tasks, environment interaction, repository-level work, and agentic tool use. |
| Qwen3-VL-Embedding and Reranker January 7, 2026 Retrieval | Extended semantic search and reranking from text to mixed text, image, screenshot, and video inputs. |
| Qwen3-Omni September 21, 2025 Omni | Native end-to-end model for text, images, audio, and video, with streamed text and speech output. |
| Qwen3-VL September 22, 2025 Vision-language | Added stronger visual perception, spatial and video reasoning, longer context, and visual agent behavior. |
| Qwen3-Max September 24, 2025 General / hosted flagship | The proprietary trillion-parameter flagship launched in an Instruct version, with a separate Thinking version in development. |
| Qwen3-Next September 10, 2025 General / architecture research | Introduced a highly sparse hybrid-attention architecture designed for faster long-context inference. Its design later informed Qwen3.5. |
| Qwen-Image August 4, 2025 Image generation | Dedicated open-weight model for text-to-image generation and image editing, with particular attention to text inside images. |
| Qwen3-Coder and Qwen Code July 22, 2025 Coder / developer tool | Paired an agent-oriented coding model with an open-source command-line coding agent. |
| Qwen VLo Preview June 26, 2025 Vision / image generation | Experimental Qwen Chat model combined visual understanding, image generation, and instruction-based editing. |
| Qwen3-Embedding and Reranker June 2025 Retrieval | Added multilingual text embeddings and reranking models for search, retrieval, clustering, and classification. |
| Qwen3 April 28, 2025 General / reasoning | Unified thinking and non-thinking modes in the main family and expanded agent, coding, multilingual, dense, and MoE options. |
| QVQ-Max March 28, 2025 Visual reasoning | Hosted visual reasoning model that analyzed images and videos before answering. |
| Qwen2.5-Omni March 27, 2025 Omni | First open Qwen model to accept text, images, audio, and video and stream both text and speech responses. |
| QwQ-32B March 6, 2025 Reasoning | Open-weight reasoning model trained with reinforcement learning for mathematics, coding, general problem solving, and tool use. |
| Qwen2.5-VL January 26, 2025 Vision-language | Improved document, chart, long-video, localization, structured-output, and visual agent capabilities. |
| QwQ-32B-Preview November 28, 2024 Reasoning | Experimental open-weight model that introduced Qwen’s dedicated text reasoning branch. |
| Qwen2.5-Coder full family November 12, 2024 Coder | Completed the coding family with six sizes from 0.5B to 32B. |
| Qwen2.5 September 19, 2024 General / Coder / Math | Released a coordinated family with better structured output, longer generation, code, mathematics, and instruction following. |
| Qwen2-VL August 29, 2024 Vision-language | Added dynamic image resolution, video understanding, visual localization, and wider multilingual text recognition. |
| Qwen2-Audio August 9, 2024 Audio-language | Accepted speech, sound, music, and text as input and returned text directly from audio. |
| Qwen2 June 7, 2024 General | Expanded multilingual coverage and improved long-context, coding, mathematics, and model-size choices. |
| CodeQwen1.5 April 16, 2024 Coder | Established Qwen’s first dedicated open coding model family. |
| Qwen1.5 February 4, 2024 General | Expanded the size range, improved developer compatibility, and later added Qwen’s first MoE release. |
| Qwen-72B and Qwen-1.8B November 30, 2023 General | Extended the first generation to both a small model and a much larger flagship. |
| Qwen-14B and Qwen-Agent September 25, 2023 General / agent framework | Added a mid-sized language model and released the framework used to build Qwen assistants with tools. |
| Qwen-VL August 22, 2023 Vision-language | First open Qwen models for image understanding, text reading, grounding, and multimodal chat. |
| Qwen-7B and Qwen-7B-Chat August 3, 2023 General | First public open-weight Qwen language models. |
| Tongyi Qianwen April 11, 2023 Hosted language model | Alibaba Cloud publicly introduced the original model and opened enterprise beta access in China. |
Current Qwen Model and Product Lineup
Qwen3.8 now spans open-weight dense and MoE models, hosted Max and Flash services, the Omni line for multimodal understanding, and LiveTranslate for real-time interpretation. Qwen3.8-Max-0902 is the current dated Max snapshot. Qwen3.8-Omni-Flash is the newest hosted Omni model, and Qwen3.8-LiveTranslate is the newest specialist speech-translation release. Qwen Studio is the user-facing app, Qwen Code is the terminal coding agent, and Model Studio and QwenCloud provide API access.
| Model or product line | What it is for |
|---|---|
| General Qwen models Foundation models | Language, reasoning, coding, tool use, and increasingly native visual understanding. Examples: Qwen3.8-27B, Qwen3.8-2.4T-A95B, Qwen3.8-Flash-Next, Qwen3.8-Flash, Qwen3.8-Max, Qwen3.8-Max-0902 |
| Qwen Coder Specialist models | Code generation, completion, repair, repository understanding, and coding agents. Examples: Qwen2.5-Coder, Qwen3-Coder, Qwen3-Coder-Next |
| Qwen VL Vision-language models | Images, documents, charts, video, visual localization, and visual reasoning. Examples: Qwen2.5-VL, Qwen3-VL |
| Qwen Omni Full multimodal models | Text, image, audio, and video understanding. Output capabilities vary by release. Some Omni models return speech. Qwen3.8-Omni-Flash returns text; the Realtime version handles live audio-video interaction. Examples: Qwen2.5-Omni, Qwen3-Omni, Qwen3.5-Omni, Qwen3.8-Omni-Flash, Qwen3.8-Omni-Flash-Realtime |
| Qwen reasoning Reasoning models or modes | Longer inference for mathematics, coding, planning, problem solving, and visual reasoning. Examples: QwQ, QVQ, Qwen3 Thinking |
| Media models Generative and speech models | Image generation and editing, speech recognition and synthesis, soundscape generation, and audio understanding. Examples: Qwen-Image-3.0, Qwen-Audio-3.1-ASR, Qwen-Audio-3.1-TTS-Next, Qwen-Audio-3.1-Realtime-Plus, Qwen3.8-LiveTranslate |
| Retrieval and safety Specialist models | Embeddings, reranking, multimodal retrieval, and content safety classification. Examples: Qwen3-Embedding, Qwen3Guard |
| Qwen Studio User product | Web, desktop, and mobile access to hosted Qwen models and tools. Examples: Chat, files, images, agents |
| Qwen Intelligence Smartphone agent platform | A business-facing Qwen agent platform for phone makers. It is designed to carry out complex tasks across mobile apps. |
| Qwen-Agent Developer framework | Tool calling, planning, memory, code execution, retrieval, and MCP-based applications. Examples: Browser Assistant, Code Interpreter |
| Qwen Code Developer tool | Terminal-based coding agent for inspecting repositories, editing files, running commands, and completing development tasks. Examples: Qwen Code CLI |
| Model Studio Cloud platform | Hosted Qwen APIs, deployment services, model evaluation, tuning, and access to third-party models. Examples: Alibaba Cloud Model Studio |
How to Read Qwen Model Names
Qwen names combine a generation, a size or service tier, and sometimes a training or capability label. For example, Qwen3.6-35B-A3B belongs to the Qwen3.6 generation, has about 35 billion total parameters, and activates about 3 billion parameters for each token. Qwen3-VL-8B-Thinking identifies an 8B vision-language model tuned for longer visual reasoning.
Qwen is the international model name and a shortened form of Qianwen. Tongyi Qianwen is the original Chinese product name, and Qwen Studio is the user application.
| Name or suffix | Meaning |
|---|---|
| 0.5B, 7B, 32B | The approximate number of model parameters. Larger numbers usually require more memory. Parameter count alone does not determine quality. |
| A3B or A22B | The approximate number of active parameters in a mixture-of-experts model. Qwen3-30B-A3B has about 30B total parameters and activates about 3B for each token. |
| Base | A pretrained checkpoint intended for research, adaptation, or further training. Chat use normally requires a post-trained variant. |
| Chat or Instruct | A post-trained model designed to follow instructions and respond conversationally. |
| Thinking | A model or mode that spends more inference time on multi-step reasoning before giving its final answer. |
| VL | Vision-language. These models understand images, documents, charts, and video alongside text. |
| Omni | A full multimodal line for text, image, audio, and video. Output capabilities vary by model. Some Omni releases return speech. Qwen3.8-Omni-Flash returns text only. |
| Coder, Math, ASR, TTS, LiveTranslate | Specialist lines for programming, mathematics, speech recognition, speech generation, and real-time multilingual interpretation. |
| Plus, Max, Flash, Turbo | In hosted model names, these labels often describe service tiers: Max usually identifies a flagship, Plus a general production model, and Flash or Turbo a faster service. They do not guarantee downloadable weights. |
| Preview or a dated suffix | An early model or a fixed service snapshot. Its announcement date, API availability date, and later stable release can differ. |
Qwen Timeline by Year
2026: Qwen3.5 Through Qwen3.8 Add Multimodal Agents and Media Models
| Release | What changed |
|---|---|
| Qwen-Image-2.1-Turbo and image APIs October 9 | The accelerated open-weight checkpoint brings eight-step generation and editing at 2K resolution to the Qwen-Image-2.1 architecture. Pro and Turbo hosted APIs are available. The downloadable weights use the Qwen Research License. |
| Qwen4 training and Qwen-Image 3.1 roadmap September 22 | Alibaba said Qwen4 is in training and Qwen-Image 3.1 is planned for later in 2026. It gave no public release date for Qwen4. |
| Qwen Intelligence September 22 | Alibaba announced a Qwen-powered smartphone agent platform for phone makers, designed to handle complex tasks across apps. |
| Qwen-Audio-3.1 and TTS-Next September 22 | The updated speech suite includes refreshed ASR and TTS models. TTS-Next can generate speech, sound effects, and ambient audio together. |
| Qwen3.8-Omni-Flash-Realtime September 21 | The hosted model supports live audio-video interaction with text and audio output, multichannel audio, and remote MCP tools. |
| Qwen-Image-2.1 September 20 | The unified image-generation and editing model adds native RGBA transparency, subject extraction, and up to 10 reference images. Its visual generation component has 7B parameters. Commercial use of the weights requires a separate license. |
| Qwen-Audio-3.1-Realtime-Plus September 20 | The duplex speech model adds eight new system voices, a 262,144-token context window, function calling, web search, and voice cloning. |
| Qwen3.8-Omni-Flash September 18 | The hosted Omni release accepts text, images, audio, and video and returns text. It adds adjustable thinking, function calling, web search, a one-million-token context window, caching, and multichannel spatial-audio input for audio-video analysis and agent workflows. |
| Qwen3.8-LiveTranslate September 18 | The real-time interpretation model separates speakers, keeps source and translated text synchronized, uses earlier context to resolve names and terminology, and can preserve speaker timbre in translated speech. The API model recognizes 60 source languages and supports speech output in 29 languages. |
| Qwen-Drive-1.0 September 3 | The autonomous-driving research model builds on Qwen3.5-4B and joins vision-language understanding with 3D perception and motion planning through external BEV and planning modules. |
| Qwen3.8-Max-0902 September 2 | The upgraded hosted Max snapshot improved engineering-scale coding, long-horizon autonomous development, multi-tool collaboration, and visual understanding. It retains a one-million-token context window and built-in tools. |
| Qwen3.8-Flash-Next August 26 | The latest open-weight Qwen3.8 model introduced Qwen Sparse Attention, Gated Residual, n-gram embedding, and a revised training recipe. The multimodal MoE acts as an early architecture preview for Qwen4. Qwen3.8-Flash is its production counterpart, with a one-million-token context window and built-in tools. |
| Qwen3.8-27B August 14 | The dense Apache 2.0 model combined native image and video understanding with controllable thinking, coding, research, and long-horizon agent work. It supports 262,144 tokens natively and can be extended to one million tokens. |
| Qwen3.8-2.4T-A95B August 8 | Qwen released its first downloadable Max-class model. This text-only MoE has 2.4 trillion total parameters, 95 billion active parameters, 92 layers, and a hybrid Gated DeltaNet and Gated Attention architecture. It requires thinking mode, supports low, medium, and xhigh reasoning_effort settings, and has a native 262,144-token context window that can be extended to 1,010,000 tokens. |
| Qwen3.8-Max August 2 | The 2.4-trillion-parameter hosted flagship supports autonomous coding, long-running professional work, and multimodal feedback for planning and self-correction. |
| Qwen-Audio-3.0-ASR-Flash July 31 | The speech-recognition model added stronger context consistency, domain-term recognition, custom hotwords, and structured transcript output. |
| Qwen-Audio-3.0-TTS July 23 | Flash and Plus models added expressive inline tags, natural-language style control, 16-language support, noisy-reference handling, and one-pass speech generation up to three minutes. |
| Qwen-Image-3.0 July 22 | The third-generation image model focused on rich layouts, fine text and formula rendering, 12 languages, 100-plus art styles, realistic interfaces, and live web retrieval. |
| Qwen3.8-Max-Preview July 19 | The hosted preview became available through Token Plan, Qoder, and QoderWork. Qwen3.8 is a continuously evolving 2.4-trillion-parameter system. |
| Qwen3.7-Max visual snapshot June 10 | The dated API snapshot added visual understanding to the Max line. Qwen3.7-Max added visual perception for real-world scenes and multimodal agent interactions. |
| Qwen3.7-Plus June 1 | The Plus model added vision-language capabilities to the Qwen3.7 generation. It is a multimodal agent model for visual understanding, coding, and tool-based tasks. It is available as a hosted model. Qwen did not release a matching open-weight checkpoint. |
| Qwen3.7-Max May 20 | Alibaba introduced its new proprietary flagship for the agent era. The model targets code development, office automation, tool use, and autonomous workflows that continue across long sequences of actions. |
| Qwen3.6-27B April 22 | This dense open-weight model brought Qwen3.6’s multimodal reasoning and agentic coding to a widely used local-deployment size. Its dense architecture also avoided the routing requirements of an MoE model. |
| Qwen3.6-Max-Preview April 18 | The hosted preview improved coding-agent performance, world knowledge, and instruction following over Qwen3.6-Plus. It was a proprietary preview under active development. |
| Qwen3.6 open models April | Qwen released selected Qwen3.6 weights, including an efficient 35B-A3B MoE model and a 27B dense model. The open family retained text and visual reasoning and placed more emphasis on real repository work. |
| Qwen3.6-Plus April 2 | The hosted Plus model combined a one-million-token context window with stronger repository-level coding, visual coding, document analysis, video reasoning, and multi-step tool work. It was released through Qwen’s user product and Alibaba Cloud Model Studio. |
| Qwen3.5 February 15 | Qwen3.5 made native multimodality part of the main open-weight family. The first release, Qwen3.5-397B-A17B, combined text, image, and video understanding with reasoning, coding, and agent abilities. |
| Qwen3-Coder-Next February 2 | The open-weight coding model built on Qwen3-Next-80B-A3B-Base focused on coding agents, local development, repository-level work, executable tasks, and environment interaction. |
| Qwen3-ASR and Qwen3-TTS January | Dedicated open speech families separated automatic speech recognition and speech generation from the audio and Omni models. ASR covered multilingual transcription and alignment. TTS focused on expressive and streaming speech synthesis. |
| Qwen3-VL-Embedding and Reranker January 7 | The retrieval line gained a shared representation for text, images, screenshots, documents, and video. This supported cross-modal search and ranking. |
2025: Reasoning, Agents, Coder, VL, Omni, and Image Generation
| Release | What changed |
|---|---|
| Qwen3-Max September 24 | The proprietary Max line scaled the Qwen3 architecture beyond the downloadable flagship models. The initial hosted release centered on instruction following, coding, agents, and general knowledge, with a separate Thinking version developed for heavier reasoning. |
| Qwen3-VL September 22 | The third VL generation added stronger image and video reasoning, spatial understanding, visual agent behavior, and long-context support. Instruct and Thinking variants separated direct visual responses from longer visual reasoning. |
| Qwen3-Omni September 21 | The native Omni model handled text, images, audio, and video in one end-to-end system. It could return text and speech in real time within one end-to-end system. |
| Qwen3-Next September 10 | This experimental architecture combined hybrid attention, greater MoE sparsity, and multi-token prediction. It reduced the cost of long-context inference and became an architectural step toward Qwen3.5. |
| Qwen-Image August 4 | Qwen released a dedicated open-weight image model after earlier visual-generation experiments in Qwen Chat. The model supported generation and editing and paid particular attention to Chinese and English text rendering. |
| Qwen3-Coder and Qwen Code July 22 | Qwen3-Coder was trained for repository-scale coding, browser use, tool calls, and multi-step software tasks. Qwen Code added a terminal agent that could apply those capabilities to files and commands. |
| Qwen VLo Preview June 26 | VLo connected image understanding with image creation and editing inside Qwen Chat. It was a hosted preview that preceded the later open-weight Qwen-Image release. |
| Qwen3-Embedding and Reranker June | These models added multilingual vector representations and relevance scoring for search and retrieval systems. They served a different role from Qwen’s generative chat models. |
| Qwen3 April 28 | Qwen brought deliberate reasoning into its main model generation. Most initial models could switch between thinking and non-thinking behavior. The family also expanded language coverage, agent use, and efficient MoE deployment. |
| QVQ-Max March 28 | The visual reasoning line moved beyond ordinary image description. QVQ-Max inspected images and videos and used longer reasoning to solve visual mathematics, charts, code, and practical questions. |
| Qwen2.5-Omni March 27 | The first open Omni model accepted text, images, audio, and video and returned streamed text or speech. Its Thinker-Talker design joined perception and spoken response within one end-to-end model. |
| QwQ-32B March 6 | The production-oriented open reasoning release used reinforcement learning for mathematics, code, general problem solving, and tools. It replaced the earlier preview as the clearest separate Qwen reasoning model before Qwen3. |
| Qwen2.5-VL January 26 | The updated VL family improved document parsing, charts, long video, object localization, and structured output. It could also act as a visual agent that selected actions from screen content. |
2024: Qwen1.5, Qwen2, Qwen2.5, and the First Reasoning Models
| Release | What changed |
|---|---|
| QwQ-32B-Preview November 28 | QwQ, short for Qwen with Questions, was Qwen’s first public model built around extended text reasoning. The experimental release exposed both the promise and early problems of long reasoning, including loops and mixed-language output. |
| Qwen2.5-Coder full family November 12 | The Coder line expanded to six sizes from 0.5B through 32B. Developers could choose between compact local models and a larger model for code generation, repair, completion, and code-agent research. |
| Qwen2.5 September 19 | Qwen2.5 arrived with general language, Coder, and Math models under one generation. It improved structured data, JSON output, long responses, instruction following, coding, mathematics, and the range of available model sizes. |
| Qwen2-VL August 29 | The second visual family processed images at dynamic resolutions and added video understanding and visual localization. Smaller models were open under Apache 2.0. The largest model initially appeared through the API. |
| Qwen2-Audio August 9 | The new audio-language model interpreted speech, environmental sounds, and music from direct audio input. It also accepted spoken instructions directly from audio input. |
| Qwen2 June 7 | Qwen2 improved coding and mathematics, supported many more languages, and extended selected models to 128K context. The family included both dense models and an MoE option. |
| CodeQwen1.5 April 16 | The first dedicated Qwen coding family separated code training from the general chat line. It included base and chat checkpoints and laid the groundwork for the later Qwen-Coder name. |
| Qwen1.5 February 4 | Qwen1.5 expanded the model range and reduced setup friction through direct integration with Hugging Face Transformers. A later MoE checkpoint introduced Qwen’s first public mixture-of-experts model. |
| Qwen-VL-Plus and Qwen-VL-Max January 25 | Hosted upgrades improved high-resolution image reading, document analysis, and visual reasoning. They extended the first VL generation before Qwen2-VL arrived later that year. |
2023: Tongyi Qianwen Becomes an Open Model Family
| Release | What changed |
|---|---|
| Qwen-72B and Qwen-1.8B November 30 | The first generation expanded in both directions. Qwen-72B became its large open flagship. Qwen-1.8B supplied a compact option for hardware with tighter memory limits. |
| Qwen-14B and Qwen-Agent September 25 | Qwen added a model between 7B and 72B and released its application framework. Qwen-Agent supported tool use and became the base for examples such as a browser assistant and code interpreter. |
| Qwen-VL and Qwen-VL-Chat August 22 | The first visual branch accepted images and text and returned natural-language answers or locations inside an image. It supported document text, multiple images, and visual grounding in Chinese and English. |
| Qwen-7B and Qwen-7B-Chat August 3 | These were the first publicly released Qwen weights. The base checkpoint served research and adaptation. The Chat version was tuned for instructions and conversation. |
| Tongyi Qianwen unveiled April 11 | Alibaba Cloud introduced the original hosted large language model and opened enterprise beta applications in China. |
Qwen Model Families Explained
General Qwen Models
The general line began with Qwen-7B, Qwen-14B, and Qwen-72B. Qwen1.5 improved model choice and compatibility. Qwen2 added languages, longer context, and stronger code and math abilities. Qwen2.5 turned the release into a coordinated family of general and specialist models.
Qwen3 changed the role of the main line because reasoning no longer required a separate QwQ model. Initial Qwen3 checkpoints could use a thinking mode for difficult work or answer directly when extra reasoning was unnecessary. Qwen3-Next then tested a more efficient hybrid architecture, which became part of the foundation for Qwen3.5.
Qwen3.5 made image and video understanding native to the main family. Qwen3.6 placed more emphasis on multimodal agents and software work, and Qwen3.7 continued that direction with proprietary Max and Plus models. Qwen3.8 then split across several deployment needs: Qwen3.8-2.4T-A95B opened the text-only Max-class foundation model, Qwen3.8-27B provided a smaller dense multimodal checkpoint, and Qwen3.8-Flash-Next previewed a more efficient Qwen4 architecture. Hosted Qwen3.8-Max and Qwen3.8-Flash provide managed context and tool features. Qwen3.8-Max-0902 updates the flagship snapshot, and Qwen3.8-Omni-Flash extends the generation to audio-video understanding.
Qwen Coder Models
CodeQwen1.5 established the separate coding branch in April 2024. Qwen2.5-Coder replaced the CodeQwen name later that year and expanded the family across six sizes. These models targeted code completion, generation, debugging, repair, and many programming languages.
Qwen3-Coder moved from code output toward agentic software work. Its training and long context supported repository-scale tasks, tool calls, browser use, and repeated interaction with an execution environment. Qwen3-Coder-Next followed in February 2026 as an open-weight model built on Qwen3-Next-80B-A3B-Base for coding agents, local development, and executable software tasks. Qwen Code supplies the terminal interface for compatible models.
The general Qwen3.6, Qwen3.7, and Qwen3.8 models also emphasize agentic coding. The specialist line uses Qwen3-Coder names, including Qwen3-Coder-Next. Repository work and software agents are also core capabilities in the general flagship models.
Qwen VL and Qwen Omni Models
Qwen VL means vision-language. These models accept images or video alongside text and produce text. The branch progressed from image description, text reading, and grounding in Qwen-VL to dynamic image resolution and video understanding in Qwen2-VL. Qwen2.5-VL strengthened document extraction, visual localization, long-video analysis, and computer use. Qwen3-VL added separate Instruct and Thinking behavior for visual tasks.
Qwen Omni covers text, images, audio, and video within one model family. Output capabilities differ by release. Qwen2.5-Omni, Qwen3-Omni, and Qwen3.5-Omni include speech-output variants. Qwen3.8-Omni-Flash focuses on audio-video understanding and text generation. It supports a one-million-token context window, thinking, function calling, web search, and multichannel spatial audio. Qwen VL focuses on image, document, and video understanding alongside text.
Qwen-Image is a separate image-generation and editing line. Qwen-Image-3.0 focuses on complex layouts, precise text rendering, multilingual output, and realistic interfaces. The later Qwen-Image-2.1 release offers a compact architecture with native transparency and multi-image editing. Qwen-Image-2.1-Turbo accelerates that checkpoint to eight denoising steps. Both 2.1 checkpoints have downloadable weights; Pro and Turbo also have hosted APIs. A VL model explains what appears in an image; an image-generation model produces a new image. Qwen VLo briefly joined both roles in a hosted preview before Qwen-Image established its own line.
Qwen Reasoning Models
QwQ was the first dedicated Qwen reasoning family. QwQ-32B-Preview demonstrated longer internal problem solving in November 2024. The March 2025 QwQ-32B release used reinforcement learning to improve mathematics, coding, tool use, and general reasoning.
QVQ applied extended reasoning to visual evidence. QVQ-Max could study an image or video and reason about diagrams, visual mathematics, code screenshots, and real-world scenes. It was a hosted model dedicated to visual reasoning.
Qwen3 absorbed reasoning into the main family through thinking modes and dedicated Thinking variants. QwQ and QVQ document the earlier development of text and visual reasoning in Qwen. Later general Qwen models provide deliberate reasoning directly.
Specialized Qwen Models
Qwen-Math models focus on mathematical problem solving and methods that combine written reasoning with programs or tools. Qwen-Embedding converts content into vectors for semantic search. Qwen-Reranker scores the relevance of retrieved results. The later VL versions extend those retrieval jobs to images and video.
Qwen-Audio and Qwen2-Audio understand speech, music, and other sound. The Qwen-Audio-3.1 suite adds updated speech recognition and synthesis models. Qwen-Audio-3.1-TTS-Next can generate dialogue, sound effects, and ambient audio together, while Qwen-Audio-3.1-Realtime-Plus supports duplex voice conversations. Qwen3.8-Omni-Flash-Realtime handles live audio-video interaction. Qwen3.8-LiveTranslate provides multilingual interpretation with speaker separation, synchronized source and translated text, context-aware disambiguation, and translated speech that preserves speaker timbre. Qwen-Drive-1.0 applies a Qwen3.5 vision-language base to 3D driving perception and motion planning. Qwen3Guard covers content-safety classification. These releases serve narrower tasks than the main conversational line.
Major Qwen Products
Qwen Studio and Qwen Chat
Qwen’s consumer-facing chat service provides web, desktop, and mobile access to hosted models. It has carried the Qwen Chat name and is now presented within Qwen Studio across web, mobile, and desktop platforms. Available models and features can change independently from open-weight releases.
QVQ-Max, Qwen VLo, and some Max models appeared in Qwen Studio only as hosted models. Their Qwen Studio availability did not include downloadable weights.
Qwen Intelligence
Alibaba announced Qwen Intelligence on September 22, 2026, as a Qwen-powered agent platform for phone makers. The service targets complex tasks that require actions across multiple apps.
Qwen-Agent
Qwen-Agent is an open-source framework for building applications around Qwen models. It provides patterns for function calling, planning, memory, retrieval-augmented generation, code execution, browser assistance, and MCP tools. It is also a backend for parts of Qwen’s hosted chat experience.
Qwen Code
Qwen Code is an open-source command-line coding agent introduced with Qwen3-Coder in July 2025. It can inspect a project, read and edit files, run shell commands, use tools, and work through multi-step development tasks. Qwen Code is the client application. Qwen3-Coder and later compatible models provide the underlying generation and reasoning. See our Qwen Code CLI guide for installation and everyday use.
Alibaba Cloud Model Studio
Model Studio is Alibaba Cloud’s platform for accessing Qwen and other models through APIs. It supports OpenAI-compatible interfaces, multimodal inputs, application development, evaluation, tuning, and deployment services. Hosted names such as Qwen-Plus, Qwen-Max, and dated API snapshots identify service tiers or API versions. They do not imply a downloadable checkpoint with identical specifications.
How Qwen Changed Over Time
From Hosted Model to Open-Weight Family
Tongyi Qianwen began as a hosted Alibaba Cloud model in April 2023. Researchers and developers could download Qwen-7B and Qwen-7B-Chat weights four months later. Qwen-VL, Qwen-14B, Qwen-Agent, Qwen-72B, and Qwen-1.8B then expanded Qwen into a model family.
Qwen2.5 Brought the Specialist Lines Together
Qwen2 strengthened the shared language foundation. Qwen2.5 coordinated general, Coder, and Math releases. The combined range offered a clearer choice of model sizes and specialties. Developers could select models by task and hardware across the coordinated family.
Qwen VL Moved Beyond Image Recognition
The first Qwen-VL recognized and located content in images. Qwen2-VL added stronger video and resolution handling. Qwen2.5-VL parsed documents and operated as a visual agent. QVQ and Qwen3-VL added longer visual reasoning. By Qwen3.7-Plus, visual perception had become part of the hosted general agent line.
Reasoning Moved into the Main Model Line
QwQ began as a separate experiment because extended reasoning required different training and response behavior. QwQ-32B followed as a downloadable model for mathematics, coding, and tools. QVQ extended the approach to visual evidence. Qwen3 then combined thinking and direct answers within the primary generation. Reasoning became part of the main Qwen model line.
Qwen3.8-Flash-Next Previewed the Qwen4 Architecture
Qwen3-Next tested the hybrid architecture later used across Qwen3.5 through Qwen3.8. Qwen3.8-Flash-Next plays a similar research role for Qwen4. It replaces full global attention with Qwen Sparse Attention over selected micro-blocks, adds gated residual branches for controlled information flow, and uses n-gram embedding as a low-compute source of model capacity. The release keeps native multimodal input and long-context support and reduces the number of active parameters per token. At Alibaba’s September 22, 2026, Apsara Conference, the company said Qwen4 was in training but gave no public release date.
Qwen Code Connected the Models to Software Work
Earlier Qwen Coder models generated and repaired code. Qwen3-Coder was trained for longer interaction with repositories, browsers, tools, and execution feedback. Qwen Code placed those capabilities in a terminal agent. Qwen3.6 and Qwen3.7 later extended planning and action into the general flagship models. Qwen3.8-Max pushes the direction further with multi-day autonomous development, long-running optimization loops, and multimodal feedback during execution.
Qwen Timeline FAQs
When was Qwen first released?
Alibaba Cloud publicly unveiled Tongyi Qianwen on April 11, 2023. The first downloadable Qwen models, Qwen-7B and Qwen-7B-Chat, followed on August 3, 2023.
What does Qwen mean?
Qwen is a shortened form of Qianwen in the Chinese name Tongyi Qianwen. Qianwen means “thousands of prompts.” The name now covers Alibaba’s language, multimodal, coding, reasoning, media, retrieval, and agent model families.
What was the first open Qwen model?
Qwen-7B and Qwen-7B-Chat were the first public open-weight Qwen language models. Alibaba released them through ModelScope and Hugging Face on August 3, 2023.
What is the latest Qwen model?
As of October 9, 2026, the newest release in this timeline is Qwen-Image-2.1-Turbo, an open-weight image-generation and editing checkpoint that uses eight denoising steps. It follows Qwen-Image-2.1, released on September 20. Pro and Turbo hosted image APIs are also available. For general-purpose models, Qwen3.8-Max is the hosted flagship and Qwen3.8-Flash-Next is the latest open-weight general model.
Has Qwen4 been released?
No. Alibaba said Qwen4 was in training on September 22, 2026, but has not announced a public release date. Qwen3.8-Flash-Next is a Qwen3.8 release that previews architecture planned for Qwen4. Alibaba also described future Qwen4.5 and Qwen5 model plans with a projected scale of 5 to 10 trillion parameters.
Is there a Qwen3.8-Coder model?
No Qwen3.8-Coder model had been released as of September 24, 2026. Qwen’s specialist coding line uses Qwen3-Coder names, including Qwen3-Coder-Next. Qwen3.8-Max, Qwen3.8-Flash, Qwen3.8-27B, and other general Qwen3.8 models also include coding capabilities.
What is the difference between Qwen, Qwen Coder, and Qwen VL?
Qwen is the general foundation model line. Qwen Coder models receive additional code-focused training for programming and software agents. Qwen VL models process images, documents, and video alongside text. Recent general models overlap with both specialist branches. The Coder and VL names identify their primary training focus.
Is Qwen a reasoning model?
Some Qwen models are designed specifically for reasoning, including QwQ and QVQ. Qwen3 and later general generations also provide thinking modes or Thinking variants. Reasoning is now part of the main Qwen line.
What is the difference between QwQ and Qwen3?
QwQ is a dedicated text reasoning family based on the Qwen2.5 era. Qwen3 is a later general-purpose generation that combines reasoning, direct responses, multilingual language work, coding, and agent capabilities. QwQ established the approach; Qwen3 brought it into the main family.
Is Qwen open source?
Many Qwen models have downloadable weights and open-source code. The license depends on the model. Most Qwen2 and Qwen3 open models use Apache 2.0, and Qwen3.8-27B uses Apache 2.0. Qwen3.8-Flash-Next uses the Qwen Community License 1.0. Qwen3.8-2.4T-A95B uses the Qwen3.8-Max License. Qwen-Image-2.1 and Qwen-Image-2.1-Turbo use the Qwen Research License, which permits noncommercial research and evaluation and requires a separate license for commercial use. These Qwen-specific licenses include conditions beyond Apache 2.0. Check the license packaged with each checkpoint before deployment. Plus, Flash, Max, and Qwen3.8-Omni-Flash are hosted services. Qwen includes both open-weight and hosted models.
What are Qwen Studio, Qwen Code, and Qwen-Agent?
Qwen Studio is the user-facing service for chatting with hosted Qwen models. Qwen Intelligence is a business-facing mobile agent platform for device makers. Qwen Code is a terminal coding agent that works with files, commands, and code repositories. Qwen-Agent is a development framework for building tool-using applications with Qwen models.
Who owns and develops Qwen?
Qwen is developed by Alibaba’s Qwen team. The project began under Alibaba Cloud. Alibaba Cloud provides many of Qwen’s hosted models and APIs.
Where can Qwen models be accessed?
Hosted models are available through Qwen Studio and Alibaba Cloud Model Studio. Open-weight releases are commonly available through Hugging Face and ModelScope, with source code and usage instructions in the QwenLM organization on GitHub.
Official Qwen resources: Qwen and Qwen Studio for hosted access; Alibaba Cloud Model Studio and QwenCloud for APIs; QwenLM on GitHub, Hugging Face, and ModelScope for code and downloadable models. Recent release pages and API references: Qwen-Image-2.1 and Turbo, Qwen-Image-2.1-Turbo weights, Qwen Research License, October 9 Turbo and API announcement, Qwen3.8-Omni-Flash, Qwen3.8-LiveTranslate, Qwen-Drive-1.0, Qwen3-Coder-Next, Qwen3.8-Flash-Next, Qwen3.8-27B, Qwen3.8-2.4T-A95B, Qwen3.8-Max, Qwen3.8-Flash, Qwen-Image-3.0, Qwen-Audio-3.0-TTS, and Qwen-Audio-3.0-ASR-Flash streaming API. Recent updates: Alibaba’s September 2026 Qwen announcements, Qwen-Audio-3.1-TTS-Next documentation, and Model Studio model lifecycle updates.
Why do some Qwen release dates differ?
Qwen releases can have different dates for an announcement, model weights, API endpoints, technical reports, and repository updates. Qwen3, for example, was announced on April 28, 2025. Its public repository release is dated April 29.
Related Resources
- Qwen Code CLI: Open-Source Command-Line AI Agent
- Qwen Code Commands Cheat Sheet: CLI, Slash Commands & Config
- Qwen API Pricing
- 10 Best CLI AI Coding Agents
- DeepSeek Timeline
- Anthropic Claude Timeline
- OpenAI and ChatGPT Timeline
- Meta AI Timeline
