Qwen 4 Release Date: Latest News, Rumors, and Expected Features

TL;DR

• Qwen 4 hasn't been released yet. As of November 7, 2026, Alibaba hasn't published an official release date.
• A release in November 2026 looks increasingly likely. An attendee at Alibaba's October 29 conference reported that the CTO had mentioned a launch the following week.
• Alibaba confirmed Qwen 4 was in training on September 22, 2026. The company also outlined plans for Qwen 4.5 and Qwen 5, with future models potentially reaching 5–10 trillion parameters.
• Qwen 4's architecture has already been previewed. Qwen3.8-Flash-Next, released in August, introduces new attention, residual, and embedding mechanisms intended for Qwen 4.
• A 1-million-token context window is possible. The experimental architecture supports context extension to 1 million tokens, and Qwen Code has added preliminary support for Qwen 4 models with that window size.
• Official specifications, benchmarks, and API prices are still unknown. There are rumors about Qwen 4 Max, Plus, Flash, and smaller open-weight models, but Alibaba hasn't confirmed the final lineup.

When Will Qwen 4 Be Released?

Our current prediction is November 2026, potentially within the first half of the month. This isn't an official release date, but it's supported by several recent developments.

The first confirmation came on September 22, 2026, when Alibaba announced at its Apsara Conference that Qwen 4 was already in training.

The company didn't give a release window, but it also announced plans for Qwen 4.5 and Qwen 5.

The next important clue appeared on October 29 during the Apsara Conference in Amsterdam.

John T Davies, who attended the conference, posted on X that Alibaba Cloud's CTO had mentioned a launch the following week. His post also referenced a presentation by Qwen research scientist Xudong Guo.

Counting from October 29, that would put the release somewhere between November 2 and November 8.

However, Alibaba's official announcement from the Amsterdam conference doesn't include this date. The claim comes from an attendee's account of a statement made during the event, rather than a written announcement from Alibaba.

The timeline became more interesting on November 6, when Qwen's official X account published a teaser:

The accompanying image featured a Chinese literary reference, 北冥有鱼，其名为_, with the final name deliberately left blank. The quotation comes from the ancient Chinese text Zhuangzi and introduces a mythical fish named Kun, which transforms into a giant bird.

The post doesn't explicitly mention Qwen 4, but its timing, imagery, and wording have fueled speculation about a new model release.

With the first week of November nearly over, the original rumored window may prove inaccurate. Still, the official teaser suggests Alibaba is preparing a new announcement.

Qwen 4 release timeline

What Is Qwen 4?

Qwen 4 is the upcoming fourth major generation of Alibaba's Qwen family of large language models. Like its predecessors, it's being developed by Alibaba's Qwen team, which has become one of the most prominent developers of open-weight AI models.

The original Qwen 3 launched in April 2025, introducing hybrid thinking modes and a collection of dense and Mixture-of-Experts (MoE) models.

Since then, Alibaba has continued improving the third generation rather than immediately moving to Qwen 4.

In February 2026, it released Qwen 3.5, introducing native multimodal capabilities and a more efficient architecture. Further releases culminated in Qwen3.8-Max, a 2.4-trillion-parameter model with 95 billion active parameters and a context window of up to 1 million tokens.

Qwen 4 will build on this work, although it won't simply be a larger version of Qwen3.8.

The Qwen team has already introduced several changes to the underlying architecture, suggesting that efficiency and long-context processing will be major priorities for the new generation.

What We Already Know About Qwen 4's Architecture

Usually, the architecture of a new AI model remains secret until launch, but in this case Alibaba was quite liberal with giving away details about the new model, so we already know a lot about it (relatively speaking).

In August 2026, the Qwen team released Qwen3.8-Flash-Next on Hugging Face, explicitly describing it as an experimental preview of the architecture that will underpin Qwen 4.

This is important because Qwen3.8-Flash-Next isn't just another small model. It introduces several technical changes that Alibaba intends to use in the next generation.

Here are its specifications:

These specifications belong to the experimental model, not necessarily to the final Qwen 4 release.

Still, the official GitHub repository describes several changes that are particularly relevant.

1. More efficient long-context attention

One of the major changes is the combination of Gated DeltaNet with Qwen Sparse Attention (QSA).

Traditional attention becomes increasingly expensive as the context grows because the model must account for relationships between more tokens.

Qwen Sparse Attention reduces this work by using a lightweight indexing mechanism to select relevant portions of the context rather than processing everything equally.

It works at the micro-block level, grouping nearby tokens during selection.

For Qwen 4, this could make long-context tasks such as analyzing large codebases or processing extensive document collections faster and less expensive.

2. Gated Residual connections

Qwen3.8-Flash-Next also introduces Gated Residual connections.

Instead of using a single residual stream to carry information between layers, the architecture expands it into four branches. Dynamic gates control how information flows through them.

The purpose is to improve information sharing across layers and make training more stable as models grow.

This is particularly relevant for Qwen 4 because Alibaba is working toward much larger model architectures.

3. N-gram embeddings

Another interesting addition is N-gram Embedding.

The experimental model uses an additional 51 billion embedding parameters alongside its 125-billion-parameter main model. These embeddings capture local token patterns and can be stored separately from the main model's GPU memory.

This allows Alibaba to increase the model's effective capacity without increasing computation proportionally.

For developers running smaller MoE models, this approach could eventually make high-capacity models more affordable to serve.

4. Lower training and inference costs

Perhaps the most interesting result is efficiency.

According to the Qwen team, training Qwen3.8-Flash-Next costs approximately one-ninth as much as training Qwen3.7-Plus, while delivering better results on coding and office tasks.

That's a substantial reduction, although the comparison concerns these two particular models rather than the final Qwen 4 family.

It also explains why Alibaba is experimenting with new architectures instead of relying entirely on larger parameter counts.

Qwen 4 Model Lineup: Max, Plus, Flash, and Open-Weight Models

Several rumored Qwen 4 variants have circulated online, including Qwen 4 Max, Qwen 4 Plus, Qwen 4 Flash, and Qwen 4 27B.

Alibaba hasn't confirmed this lineup.

However, the existing Qwen naming system gives us some idea of what the next family might look like.

There's another small clue in the Qwen Code repository. The October 28 update includes test identifiers such as qwen4-max, qwen4-plus, and qwen4-coder-plus.

But these are examples used to test how the software handles model names. They don't establish that Alibaba will release models with those exact names.

Will Qwen 4 be open source?

Alibaba has a strong history of releasing downloadable model weights, so it's reasonable to expect at least some open-weight Qwen 4 models.

The company went especially far with Qwen3.8-Max, releasing the weights of a flagship model with 2.4 trillion parameters.

Still, Qwen 4's distribution strategy remains unknown.

Alibaba may release smaller models immediately and larger models later, or use a combination of hosted and open-weight releases.

The license is also worth watching. Qwen3.8-Flash-Next uses the qwen-community-1.0 license, but that doesn't establish what license the full Qwen 4 family will use.

How Powerful Will Qwen 4 Be?

Alibaba hasn't published Qwen 4 benchmark results, so there's no reliable way to compare it with GPT-6, Claude, Gemini, or other frontier models yet.

What we can examine is the direction in which Qwen's capabilities have been developing.

The August release of Qwen3.8-Max made a particularly strong push toward autonomous coding and long-running agent tasks.

According to Alibaba's official technical announcement, Qwen3.8-Max achieved the following results:

Interestingly, the improvements aren't uniform. GPQA Diamond barely changed, while coding and agent benchmarks improved substantially.

That suggests Alibaba is putting considerable effort into models that can work on complex tasks over extended periods.

Qwen 4 Pricing and Availability

Alibaba hasn't announced Qwen 4 API pricing or confirmed where the models will become available first.

Based on previous Qwen launches, the most likely distribution channels are Qwen's own products and Alibaba Cloud's model services.

Qwen3.8-Max, for example, was initially released through Alibaba's hosted API, with open weights following later.

Qwen3.8-Flash-Next is already available through Hugging Face and supported by the Qwen ecosystem.

For Qwen 4, there are three likely ways to access the new models:

• Qwen's chat and cloud services: The most likely starting point for trying new hosted models.
• Alibaba Cloud Model Studio and QwenCloud: Potential API access for developers.
• Hugging Face and ModelScope: Possible distribution channels for open-weight versions.

An immediate open-weight release isn't guaranteed, and neither is free API access.

As for pricing, Alibaba has made inference efficiency a major development priority. But architectural improvements don't automatically mean lower API prices, especially if Qwen 4 offers a much larger flagship model.

We'll need the official price list before making a meaningful cost comparison.

What Comes After Qwen 4?

Alibaba has already outlined plans for Qwen 4.5 and Qwen 5.

The company's September 2026 roadmap specifically identifies these generations as candidates for models reaching 5–10 trillion parameters.

That implies Qwen 4 may be the beginning of another extended model generation, with subsequent updates improving performance and scaling the architecture.

There's already a precedent for this approach.

The Qwen 3 family launched in April 2025 and continued receiving substantial architectural and performance updates throughout 2026, rather than being replaced immediately.

If Alibaba follows a similar pattern, Qwen 4 could remain the foundation for several successive releases.

However, no dates have been announced for Qwen 4.5 or Qwen 5.

Bottom Line

Qwen 4 is officially in development, and the evidence for an imminent release is stronger than it was a few months ago.

Alibaba has confirmed that the model is being trained, published an experimental version of its architecture, and updated its tooling to handle anticipated Qwen 4 model identifiers. The October 29 conference report and November 6 teaser further suggest that a launch could be close, although Alibaba still hasn't confirmed a date.