Alibaba's Qwen 4 model remains in training despite release rumors

The official status of Qwen 4

As of November 5, 2026, Alibaba has not released the Qwen 4 model. The company's official stance, shared in a press release on September 22, 2026, is that the next-generation model is "currently in training." There is currently no model card, download link, API name, or pricing available for the public.

The current generation of models remains Qwen 3.8. The flagship of this series, Qwen3.8-Max, launched on August 3, 2026, and was followed by the release of its weights as Qwen3.8-2.4T-A95B. Users looking for the newest capabilities should continue to use the 3.8 series until Alibaba provides a formal announcement.

Despite the lack of official documentation, there is significant noise regarding a release date. Some model databases incorrectly list a June 2026 release, but these entries are not supported by any official Qwen repositories. The Qwen organization on Hugging Face currently lists its newest upload as Qwen-Image-2.1-Turbo from October 9, 2026.

Tracing the 'next week' rumors

Much of the current speculation stems from a single report from an attendee at Alibaba Cloud’s Apsara Conference in Amsterdam. On October 29, 2026, an attendee reported that Alibaba Cloud’s CTO, Feifei Li, stated "Qwen4 next week" during a presentation. This would suggest a potential release window between November 2 and November 8, 2026.

However, Alibaba’s own written account of the Amsterdam event does not mention Qwen 4. The company's press release for the event focused on its first cloud region in the Netherlands and did not include any specific details about the new model. This discrepancy suggests that the "next week" claim rests on a single person's note of a stage remark rather than a confirmed corporate schedule.

Prediction markets are reacting to these signals with high confidence. On the Manifold platform, a market for the Qwen 4 release date gave a 93.5% probability to a release before December 2026. Another market showed an 83% probability for Alibaba publishing open weights for a Qwen 4 model on Hugging Face by November 25, though these markets are small and reflect crowd sentiment.

Expected model lineup and specs

Alibaba has provided some clues regarding the naming conventions for the upcoming series. Slides shown at the Apsara Conference in Hangzhou and Amsterdam list four specific tiers: Qwen4-Max, Qwen4-Flash, Qwen4-Plus, and Qwen4-27B. While no specific sizes or prices have been published, these names suggest a tiered approach similar to the previous generation.

Internal code changes in Qwen Code, Alibaba’s open-source coding agent, provide further technical hints. Developers merged pull requests to size the Qwen 4 generation at a 1,000,000-token context window. This would be an increase from the current default of 262,144 tokens, though these tests were run against a fake server rather than a live endpoint.

The code also indicates the inclusion of reasoning controls for the Qwen 4 family. These controls allow users to toggle the "thinking" effort of a model between low, medium, and xhigh levels. This suggests that the new models may focus on complex problem-solving tasks where the model takes more time to process information before generating a final answer.

Architecture and open source plans

The design for Qwen 4 is already partially accessible through a preview model. On August 26, 2026, the team released Qwen3.8-Flash-Next, which serves as an early preview of the architecture used in Qwen 4. The team stated that this model allows the community to examine architectural changes before the full Qwen 4 family is built on top of them.

A key architectural bet for Qwen 4 involves an n-gram lookup table. Instead of spending compute on every single parameter, a large share of the parameters are parked in this table, which is cheap to read and can sit in ordinary memory. If Qwen 4 follows this design, its models may be more cost-effective to run than their total parameter counts would typically suggest.

The question of whether Qwen 4 will be open source remains unanswered as Alibaba has not announced a license. However, the Qwen 3.8 generation set a precedent by offering different licenses for different tiers. For example, the 27B model used the permissive Apache 2.0 license, while the Max model required a specific license for businesses exceeding certain revenue or user thresholds.

What to use while you wait

While waiting for the official Qwen 4 launch, the Qwen 3.8 series remains the current standard for users. Qwen3.8-Max is available via Qwen Studio and provides high-end performance for those who need a flagship experience. For those looking for a balance of power and cost, Qwen3.8-Flash is available at a lower price point for API users.

For users interested in running models locally on hardware like a Mac Studio, Qwen3.8-27B is the most viable option. It fits on a Mac with 32 GB of unified memory when using a 4-bit build. Additionally, the Flash-Next model is already available for those who want to test the upcoming Qwen 4 architecture on their own machines today.

The community should keep a close watch on three primary sources for the official launch. The Qwen organization on Hugging Face, the @Alibaba_Qwen account on X, and the official blog at qwen.ai are the primary locations where Alibaba will post the model weights and official documentation. If no updates appear by November 8, the "next week" rumor will likely have been delayed.

Technical details and specs

The Qwen 4 architecture is already being integrated into various open-source tools. The llama.cpp project, which allows models to run on standard computers and Macs, has already merged support for the Flash-Next architecture under the internal name "qwen4exp." This indicates that the engineering community is preparing for the new model's arrival by building the necessary infrastructure in advance.

The transition to Qwen 4 also signals a move toward more complex reasoning capabilities. The inclusion of "reasoning-effort" tiers in the code suggests that Alibaba is refining how models handle multi-step problems. By allowing the model to spend more time "thinking" before answering, the system can potentially achieve higher accuracy on difficult logic tasks without increasing the base parameter count of every model.

Furthermore, the move toward a 1M-token context window for the Qwen 4 series is a significant technical goal. A larger context window allows the model to process and remember much longer documents or conversations in a single prompt. While this is currently a tool setting in the Qwen Code repository, it points toward a goal of making Qwen 4 more capable of handling massive amounts of data simultaneously.

Licensing and commercial use

Understanding the licensing for Qwen 4 will be critical for businesses planning to integrate the model into their products. Based on the Qwen 3.8 release, Alibaba appears to be using a tiered licensing system. The 27B model is likely to remain under the Apache 2.0 license, which allows for broad commercial use with minimal restrictions.

Larger models, such as the Qwen4-Max, may follow the precedent set by Qwen3.8-Max. This license requires businesses with over 100 million monthly users or $20 million in monthly revenue to display the model name. Additionally, companies building API services or office assistants may need to secure a separate commercial license once they exceed $50 million in annual revenue.

Because Alibaba has not yet published the specific licenses for Qwen 4, developers should wait for the official model cards. The license will depend on which tier of the model a user chooses to implement. This distinction is important because it determines the legal requirements for commercial deployment, revenue reporting, and brand attribution for AI-powered applications.

Future roadmap and scaling

The roadmap for Alibaba's AI models extends well beyond the upcoming Qwen 4 release. During the Apsara Conference in September, the company announced plans for Qwen 4.5 and Qwen 5. These future generations are projected to scale up significantly, with parameter counts reaching between 5 and 10 trillion.

The jump to 5 to 10 trillion parameters represents a massive increase in scale compared to the current Qwen 3.8-Max, which features 2.4 trillion parameters. While these larger models are not expected to be released soon, they indicate Alibaba's long-term strategy to push the boundaries of model capacity and complexity. These figures are specifically for the generations following Qwen 4.

For now, the focus remains on the transition from Qwen 3.8 to Qwen 4. The company's strategy involves releasing architectural previews like Flash-Next to build community trust and gather feedback. This iterative approach allows the community to see the underlying technology before the final, polished models are deployed to the public and integrated into enterprise tools.

Summary of current rumors

To navigate the current information landscape, it is helpful to separate confirmed facts from speculation. The most reliable information is that Qwen 4 is in training and that the Qwen 3.8 series is the current production standard. The "next week" release date is an unverified report from a single conference attendee and should be treated with caution.

The naming of the Qwen 4 series—Max, Flash, Plus, and 27B—is the most concrete information available regarding the new lineup. These names suggest a consistent branding strategy that mirrors the previous generation. However, until Alibaba provides official specifications, these names do not come with confirmed performance benchmarks or hardware requirements.

The most reliable way to track the release is to monitor the official Qwen channels. A new repository on Hugging Face or a post on the @Alibaba_Qwen X account will be the first signs of a real release. Until then, users should continue to utilize the Qwen 3.8 models, which remain the most capable and officially supported models in the current lineup.

Hardware and local execution

The feasibility of running Qwen 4 locally will largely depend on the final parameter count of the 27B model. Based on the Qwen 3.8-27B model, which fits on a Mac with 32 GB of unified memory in a 4-bit build, it is highly likely that the Qwen4-27B will also be accessible to Mac users. The community has already begun creating 4-bit builds for the Flash-Next architecture, which serves as a preview for Qwen 4.

In contrast, the Qwen4-Max model is expected to be far too large for local execution on consumer hardware. Its predecessor, Qwen3.8-Max, has 2.4 trillion parameters, placing it firmly in the category of data-center-scale models. Users who want to use the Max-class models will likely need to access them through an API or a cloud-based application rather than downloading the weights.

The development of the llama.cpp project continues to be a key factor for local AI enthusiasts. By merging support for the Qwen 4 architecture preview, the project ensures that the community can begin testing the new technology as soon as it becomes available. This infrastructure is essential for making high-performance models like Qwen 4 accessible to developers who prefer to run models on their own hardware.

Comparison with Western flagships

The Qwen 3.8 series has been positioned as a direct competitor to Western flagship models. Alibaba's own reviews of the Qwen3.8-Max model highlight its performance in various benchmarks compared to leading models from the United States. The release of Qwen 4 will likely continue this trend of challenging the dominance of Western AI providers in the open-weights space.

One of the key advantages of the Qwen series is the availability of open weights for several tiers. This allows developers to fine-tune the models and deploy them in private environments, a capability that is often restricted with proprietary models. The Qwen 4 release will be a significant moment for the open-source community, as it provides a new set of weights to experiment with and integrate into diverse applications.

As the Qwen 4 models arrive, the community will be looking for how they stack up against the latest updates from other major AI labs. The focus will be on whether the new architecture—specifically the n-gram lookup table and the reasoning-effort tiers—provides a noticeable leap in performance. For now, the Qwen 3.8 series provides a solid baseline for comparing Alibaba's progress against the rest of the industry.

Summary of Qwen 4 rumors

To navigate the current information landscape, it is helpful to separate confirmed facts from speculation. The most reliable information is that Qwen 4 is in training and that the Qwen 3.8 series is the current production standard. The "next week" release date is an unverified report from a single conference attendee and should be treated with caution.

The naming of the Qwen 4 series—Max, Flash, Plus, and 27B—is the most concrete information available regarding the new lineup. These names suggest a consistent branding strategy that mirrors the previous generation. However, until Alibaba provides official specifications, these names do not come with confirmed performance benchmarks or hardware requirements.

The most reliable way to track the release is to monitor the official Qwen channels. A new repository on Hugging Face or a post on the @Alibaba_Qwen X account will be the first signs of a real release. Until then, users should continue to utilize the Qwen 3.8 models, which remain the most capable and officially supported models in the current lineup.

Next steps for users

For users who want to stay ahead of the curve, the best course of action is to continue using the Qwen 3.8 models while monitoring official channels. Qwen3.8-Max is the current flagship for high-performance needs, while Qwen3.8-27B is the best choice for local deployment on Apple silicon. These models are not just stopgaps; they are the current state-of-the-art for the Qwen family.

If you are a developer, you can already begin experimenting with the architecture that will power Qwen 4. By downloading and running Qwen3.8-Flash-Next, you can get a feel for the underlying design changes. This allows you to prepare your workflows and tools for the transition to the new generation before it officially launches.

Finally, keep an eye on the official Qwen blog and the Hugging Face repository. These are the primary locations where Alibaba will post the model cards, weights, and licensing details. By following these official sources, you can avoid the misinformation that often circulates on social media and prediction markets during the pre-release phase.

Summary of Qwen 4 rumors

To navigate the current information landscape, it is helpful to separate confirmed facts from speculation. The most reliable information is that Qwen 4 is in training and that the Qwen 3.8 series is the current production standard. The "next week" release date is an unverified report from a single conference attendee and should be treated with caution.

The naming of the Qwen 4 series—Max, Flash, Plus, and 27B—is the most concrete information available regarding the new lineup. These names suggest a consistent branding strategy that mirrors the previous generation. However, until Alibaba provides official specifications, these names do not come with confirmed performance benchmarks or hardware requirements.

The most reliable way to track the release is to monitor the official Qwen channels. A new repository on Hugging Face or a post on the @Alibaba_Q