Qwen is Alibaba’s model family, and it now sits under a large share of Chinese AI answers: the Qwen (千问) app, Taobao and Amap assistants, Apple Intelligence in China and many third-party products running its open weights. This page lists Qwen model, API and agent releases, newest first, with dates taken from the Qwen team’s blog, the QwenLM GitHub repositories, Hugging Face model cards and Alibaba Cloud Model Studio’s model release log. Where only the month is reliable, or the open-weight and API dates differ by a day or two, the row says so. For a plain-English introduction, read our beginner’s guide to Qwen and Qwen agents.
| Date | Update | What changed | Why it matters |
|---|---|---|---|
| 31 October 2026 | Apple confirms Qwen in Apple Intelligence China | Apple’s China newsroom and Alibaba say Qwen powers Siri hand-offs, Writing Tools and Visual Intelligence (“Ask Qwen”) from the 4 November iOS beta, with general release in iOS 27.3. | Qwen becomes an answer layer on Chinese iPhones, so more questions are answered without a click to a website. |
| 27 October 2026 | Qwen-Audio-3.1-Realtime-Flash | Model Studio adds a duplex real-time voice model with tool calling, web search and voice cloning, and a 262,144-token context. | Voice agents with live search are another route through which pages are retrieved and read aloud, not visited. |
| 13 October 2026 | Qwen Go! waitlist | The Qwen app opens sign-ups for a personal agent that books, compares prices and negotiates refunds inside Alibaba and partner services. | Task agents complete commercial journeys inside apps, bypassing the open web for some shopping and local intent. |
| 2–10 October 2026 | Qwen-Image-2.1-Pro and Turbo | Unified text-to-image and editing models with native transparency and up to 10 reference images; Turbo is listed as open-source. | Mostly a creative tool, but it speeds up e-commerce imagery on Alibaba’s marketplaces. |
| 22 September 2026 | Apsara: Qwen 4 “in training” | Alibaba says Qwen 4 is in training and projects Qwen 4.5 and Qwen 5 at 5–10 trillion parameters. It also shows Qwen Intelligence for phones, with HONOR first. | A roadmap, not a release. No Qwen 4 model card or API ID had been published by 8 November 2026. |
| 17–21 September 2026 | Qwen3.8-Omni-Flash | Model Studio adds an omni model taking text, image, audio and video input, with thinking modes, function calling and web search, plus a real-time version. | One model can now watch, listen and search, widening what an AI answer can draw on. |
| 2 September 2026 | Qwen3.8-Max-0902 snapshot | Updated Max snapshot with stronger coding, multi-tool agent work and chart and document reading, keeping a 1M-token context. | Snapshot IDs change behaviour under a fixed product name, so record the exact ID when testing AI visibility. |
| 1 September 2026 | Qwen3.7 text embedding and rerank | New embedding and rerank models; Alibaba claims about 35% better web-search reranking than qwen3-rerank. | Rerankers decide which retrieved passages reach the model, which is the step closest to ranking. |
| 26 August 2026 | Qwen3.8-Flash and Qwen3.8-Flash-Next | Hosted multimodal Flash model with a 1M-token context and OpenAI- and Anthropic-compatible APIs; Flash-Next weights published under a custom licence. | Flash is the cheap default many apps call, so it is the model most likely to summarise your pages. |
| 14 August 2026 | Qwen3.8-27B open weights | Dense multimodal 27B model with thinking and non-thinking modes, released under Apache 2.0. | A permissive, self-hostable model that third-party search and answer products can adopt freely. |
| 12 August 2026 | Qwen3.8-2.4T-A95B open weights | Open release of the Max-class model: 2.4T-parameter MoE, about 95B active, 1M-token context, under a Qwen3.8-Max licence with display and revenue conditions. | The first Max-class Qwen anyone can download, though large commercial users need a separate licence. |
| 2–3 August 2026 | Qwen3.8-Max | Hosted 2.4T-parameter MoE flagship for long-horizon coding and office work with native vision. | Alibaba’s top model for agents that research and act over long tasks. |
| 15–21 July 2026 | Qwen3.7-Flash | Multimodal Flash model with stronger search-agent and code-interpreter agent abilities. | Search agents on the cheaper tier mean more automated retrieval of web pages. |
| May–June 2026 | Qwen3.7-Max and Qwen3.7-Plus | Preview on 17 May, Max on 20–21 May (text first, vision added in the 8 June snapshot), and Plus on 26 May with GUI and screen-operation abilities. Both hosted only. | GUI agents can navigate sites and apps as a user would, not just read HTML. |
| 16–24 April 2026 | Qwen3.6 open weights and Max-Preview | Qwen3.6-35B-A3B (16 April) and Qwen3.6-27B (22 April) released as open weights; Qwen3.6-Max-Preview shown as the next proprietary model. | Strong small open models spread Qwen into self-hosted tools and European providers. |
| 2 April 2026 | Qwen3.6-Plus | Hosted model with a 1M-token default context and better agentic coding and multimodal reasoning, also used in Qwen App and Wukong. | Marks Alibaba’s shift in messaging from chat to real-world agents. |
| March 2026 | Qwen3.5-Omni | Omni Plus and Flash models (15 March snapshots) added to Model Studio, including real-time versions. | Voice and video input for assistants that answer without a screen of links. |
| 16 February 2026 | Qwen3.5 | Qwen3.5-397B-A17B open weights and the hosted Qwen3.5-Plus, pitched at “visual agentic” use; medium sizes followed on 24 February and small ones on 2 March. | The base architecture for later 3.6–3.8 models, and the first generation that can operate apps. |
| February 2026 | Qwen3-Coder-Next | New coding model added to Model Studio on 19–20 February. | Coding agents increasingly fetch documentation pages, so developer docs become retrieval targets. |
| 17 November 2025 | Qwen app public beta | Alibaba relaunches its consumer assistant as the Qwen (千问) app. | Starts Qwen’s rise as a consumer answer surface in China; QuestMobile later put it at about 167m MAU. |
| 23–24 September 2025 | Qwen3-Max, Qwen3-VL and Qwen3-Omni | At Apsara 2025 Alibaba launches the hosted Qwen3-Max, after a 5 September preview, alongside Qwen3-VL and Qwen3-Omni. | The first trillion-parameter-class Qwen and a broad multimodal line-up. |
| 11 September 2025 | Qwen3-Next-80B-A3B | Ultra-sparse MoE with hybrid attention, released under Apache 2.0. | Cheap, fast inference that later Qwen generations build on. |
| August 2025 | Qwen-Image | Image generation and editing model, with editing following in mid-August. | Better text rendering in images for marketplace and ad creative. |
| 21 July–6 August 2025 | Qwen3-2507 updates and Qwen3-Coder | Separate Instruct and Thinking updates to Qwen3, a 256K context, and Qwen3-Coder-480B-A35B on 22 July. | Splitting thinking and non-thinking variants changed how developers tune cost and latency. |
| 29 April 2025 | Qwen3 | Eight open-weight models from 0.6B to 235B-A22B, with hybrid thinking modes, under Apache 2.0. | The most widely adopted open Qwen generation, behind many third-party apps. |
| 26 March 2025 | Qwen2.5-Omni-7B | An end-to-end model that sees, hears and talks. | An early sign of voice-first assistants. |
| 6 March 2025 | QwQ-32B | Reasoning model trained with large-scale reinforcement learning, released under Apache 2.0. | Showed a mid-sized open model could rival DeepSeek-R1 on reasoning. |
| January 2025 | Qwen2.5-Max and Qwen2.5-VL | Hosted MoE flagship (28 January) that Alibaba claimed beat DeepSeek-V3, alongside the Qwen2.5-VL vision line. | Alibaba’s response to DeepSeek’s rise, and the start of the 2025 price and model race. |
| November–December 2024 | Qwen2.5-Coder-32B, QwQ-Preview and QVQ | Coding model (November), the first QwQ reasoning preview (late November) and the QVQ visual reasoning preview (December). | The start of Qwen’s reasoning branch. |
| 19 September 2024 | Qwen2.5 | Models from 0.5B to 72B, plus Coder and Math variants. Most are Apache 2.0; 3B and 72B use Qwen licences. | Qwen becomes a default open model for fine-tuning worldwide. |
| August 2024 | Qwen2-VL and Qwen2-Audio | Vision-language and audio models. | Multimodal input arrives in the open line. |
| 6 June 2024 | Qwen2 | Models from 0.5B to 72B, with most sizes under Apache 2.0 and 72B under the Qianwen licence. | Big gains in multilingual quality, which drove adoption outside China. |
| 5 February 2024 | Qwen1.5 | A refreshed family from 0.5B to 72B, with an MoE-A2.7B model on 28 March. | Brought Qwen into the main Hugging Face tool chain. |
| 30 November 2023 | Qwen-72B and Qwen-1.8B | Open models trained on 3T tokens with a 32K context, plus Qwen-Audio around the same time. | The first frontier-scale open Chinese model. |
| 25 September 2023 | Qwen-14B and Qwen-Agent | Qwen-14B released, along with the Qwen-Agent framework and qwen.cpp. | Qwen-Agent is the toolkit for tool use and RAG that later powers Qwen Chat. |
| August 2023 | Qwen-7B and Qwen-VL | Qwen-7B and Qwen-7B-Chat open-sourced on 3 August, followed by the Qwen-VL vision model. | The first open-weight Qwen release. |
| April 2023 | Tongyi Qianwen unveiled | Alibaba Cloud shows its Tongyi Qianwen model at its April 2023 summit; public access followed in September 2023 after regulatory approval. | The starting point of Alibaba’s model family. |
Key model events
Qwen has grown in three phases. From 2023 to 2024 it was mainly an open-weight project. Qwen-7B in August 2023 was followed by 14B, 72B, Qwen1.5, Qwen2 and Qwen2.5, each one bigger and more multilingual, and most of them under Apache 2.0. That made Qwen the base for thousands of fine-tuned models on Hugging Face, and it is why so many products outside China quietly run on Qwen today. Qwen-Agent arrived alongside Qwen-14B in September 2023 and has grown into the framework behind Qwen Chat, with function calling, RAG, a code interpreter and MCP support.
The second phase, in 2025, was a response to DeepSeek. Alibaba launched Qwen2.5-Max on the first day of the Lunar New Year, shipped QwQ-32B as an open reasoning model, and then released Qwen3 with thinking and non-thinking modes in one model. By September 2025 the line-up had split into hosted Max and Plus tiers, open MoE models such as Qwen3-Next, and separate vision, omni, coding, speech and image branches. The consumer side was relaunched as the Qwen app in November 2025.
The third phase, in 2026, is about agents. Qwen3.5 in February was pitched at operating apps, and each later version (3.6 in April, 3.7 in May, 3.8 in August) has led on agentic coding, long tasks and GUI control. Qwen3.8 brought a Max-class model to open release for the first time, but under a custom licence. The larger open model needs a separate licence above a US$50 million revenue threshold, while Qwen3.8-27B stays on Apache 2.0. At the Apsara Conference Alibaba said Qwen 4 is in training. On the product side it has added the Qwen Go! personal agent and the Apple Intelligence role in China, first reported as approved in October (our report) and confirmed by Apple on 31 October (details).
For search, the most important pieces are the retrieval ones. Model Studio’s web search tool lets any developer switch on live search for Qwen, and whether sources are shown depends on the endpoint. The Qwen3.7 embedding and rerank models decide which passages reach the model. Flash models are the cheap default, and their thinking-mode settings can silently produce empty answers. Open weights also mean Qwen turns up in places like Ecosia’s European stack, where the host, not Alibaba, controls retrieval. Safety is part of the record too. Cohere reported that Qwen3.8-Flash-Next made a rare unrequested upload attempt in a sandbox, which is a reminder to sandbox any agent that has tools. The practical takeaways are the same as for DeepSeek: keep your Chinese pages crawlable and easy to extract, make authorship and facts explicit, and check how Qwen answers describe your brand after each new snapshot, recording the exact model ID and region. Wider coverage is in China AI and AI News.
Sources: Qwen3 blog, QwenLM/Qwen3 news log, Qwen-Agent, Qwen3.6-Plus, Qwen3.6-35B-A3B, Qwen3.8-2.4T-A95B model card and licence, Qwen3.8-27B and the Model Studio release log. Where API listing dates and open-weight dates differ, the row gives the range.
Parallel histories: DeepSeek, ERNIE and Baidu algorithm updates. Last reviewed 8 November 2026.