On September 22, 2026, at its annual technology event Apsara Conference (Yunqi Conference) held in Hangzhou, China, Alibaba announced that its next-generation AI model Qwen 4 is currently in training. It also presented a wide range of AI-related plans, including the expected scale of subsequent models, plans for a new image generation model, an in-house AI chip and plans to expand its data centers. However, Qwen 4 itself has not been released yet, and neither a release date nor pricing has been announced.
Qwen 4 is "in training"
According to Alibaba's announcement, the next-generation Qwen 4 model is currently in the training stage (the process of having an AI read large amounts of data to acquire its capabilities).
According to KoCPC, a Taiwanese IT media outlet, Liu Dayiheng, who leads development of Qwen's large language models, took the stage at the event and revealed four names in the Qwen 4 series for the first time: Max, Plus, Flash and 27B. According to the report, the positioning of each is as follows.
- Max: the top model, aiming for peak performance
- Plus and Flash: models that balance performance, efficiency and cost
- 27B: a 27-billion-parameter model intended to be deployed and run locally
The presentation slides only said "Coming Soon," and no release date, pricing or downloadable weights (the trained data) were given for any of the models. Whether the 27B model will be released as open weights (a form in which the model weights are published so anyone can download and run them) is also unclear at this point.
Qwen 4.5 and Qwen 5 to reach 5 trillion to 10 trillion parameters
For the Qwen 4.5 and Qwen 5 series that will follow Qwen 4, Alibaba outlined plans to increase parameters (a number indicating an AI's scale) to between 5 trillion and 10 trillion. According to Reuters, the current top model, Qwen3.8-Max, has 2.4 trillion parameters, so this would be two to four times that scale. Eddie Wu, CEO of Alibaba Group, explained that the aim is to enable the models to handle more complex, longer-running tasks. This, too, is a future plan with no timeline given.
Alibaba also presented results from an effort in which AI repeatedly improves itself. According to the announcement, when Qwen3.8-Max was left to improve itself fully automatically for more than a month, it went through 33 rounds of improvement, and its score on the index of Artificial Analysis, a third-party site that compares AI performance, rose from 40 to 45.
Other models announced
Besides Qwen 4, models related to audio and images were announced. The main ones listed in the announcement are as follows.
- Qwen3.8-LiveTranslate: a simultaneous interpretation model that translates as people speak. It is said to cut the delay before the translation appears by nearly 20%, from about 2.8 seconds to about 2.3 seconds
- Qwen-Audio-3.1-TTS-Next: an audio generation model said to create entire cinematic sound scenes from a script. It is intended for use in audiobooks, films, podcasts and games
- Qwen-Image 3.1: an image generation model optimized for design work and online retail marketing. It is planned for release within the year and has not been released yet
- Qwen Intelligence: a service that provides smartphone makers with an AI agent framework for handling tasks that span multiple apps
On Alibaba Cloud's model service Model Studio, Qwen3.8-Omni-Flash, a model that understands text, images, audio and video (September 17), and a version of it that can interact in real time through audio and video (September 21) were added around the time of the event.
New AI chip and data centers
T-Head, Alibaba's chip design unit, announced Zhenwu V900, a new chip used for AI training and inference. It offers three times the performance of the previous-generation Zhenwu M890 and has 216GB of memory. Mass production and sales are scheduled to begin in January to March 2027. It is also said to support large-scale server configurations connecting up to 500,000 chips.
Wu also set a goal of expanding the capacity of data centers operated by Alibaba Cloud to more than 20GW (gigawatts) worldwide by 2032. Overseas, it plans to open its first cloud sites in Turkey, Finland and the Netherlands over the next 12 months. According to Reuters, following the series of announcements, Alibaba's shares rose 5.1% in the Hong Kong market, hitting a one-month high.
CoAI's take
Much of this announcement concerns plans for "what's next." Qwen 4 is in training, and its release date, pricing and license have not been decided. The figure of 5 trillion to 10 trillion parameters for Qwen 4.5 and Qwen 5 is also only a target at this point. On the other hand, Qwen models are already used by developers worldwide through APIs and open weights, and the reported inclusion of a 27B model that can run locally in the series is a point of interest for companies that want to run AI in their own environments. The license under which the 27B model is released will determine whether it can be used for work, so it is worth checking when it comes out. The announcement once again made clear Alibaba's strategy of building not only AI models but also chips and data centers in-house.
Sources
- Alibaba Cloud's 2026 Apsara Conference: Full-Stack AI Roadmap Along with Global Market Expansion Plan (Alibaba Group)
- Alibaba deepens AI push with new chip, bigger model; shares jump 5% (Khaleej Times/Reuters)
- Qwen4 family revealed for the first time: Max, Flash, and Plus launch simultaneously across three lines, with Qwen4 27B making an appearance (KoCPC)
- Alibaba Puts Qwen4 Family Into Training; Roadmap Points to 5–10T Qwen4.5 and Qwen5 (Pandaily)
- Newly released models (Alibaba Cloud Model Studio)
