Qwen 4 is getting closer: Only one model appears set to be open

Alibaba has shown what the next generation of its Qwen AI model will look like. At the Apsara conference in Hangzhou on September 22, 2026, Qwen 4 was presented in four tiers: Max, Plus and Flash, plus a model with 27 billion parameters (27B). Parameters are the settings an AI model learns during training, and their number is often used as a measure of the model's size.

Only one model appears set to be open

According to a blogger who summarized the presentation, Max, Plus and Flash look like models available only through paid web access (API), as with Qwen3.8. Only the 27B model is positioned as an open-weights model. Open weights mean that others can download the model itself and run it locally on their own computers.

However, there is no formally confirmed plan for which Qwen 4 weights will be open. According to a technical overview, the expectation of an open 27B model is based on earlier patterns. One assessment says Alibaba kept the top tier closed in the two previous generations, 3.6 and 3.7, while the mid-range has been open.

The license for Qwen 4 has not been published. Qwen3.8-27B uses Apache 2.0, while the preview model Flash-Next uses Qwen Community License 1.0.

Preview already in August

In August 2026, according to a news article, the Qwen team released Qwen3.8-Flash-Next with open weights on Hugging Face and ModelScope. It is a multimodal model, meaning it can work with several kinds of data, and it has 125 billion parameters plus 51 billion N-gram embedding parameters. The team describes it as a preview of the architecture, not as the final model. It is not clear whether it is the same as Qwen 4 Flash, which other sources describe as an API tier.

When will it arrive?

Alibaba said at the conference that Qwen 4 was in the training phase. No release date, price, context window or test results were given.

There are rumors, though. An attendee at Apsara in Amsterdam on October 29 reportedly heard chief technology officer Feifei Li say "next week," which would point to November 2-8. This is unconfirmed, and no official Alibaba channels have repeated it. In addition, a code change in the Qwen Code repository on October 28 gives the Qwen 4 family a context window of 1 million tokens, meaning how much text the model can handle at once. The author himself writes that the figure was passed along and is not officially confirmed.

Bigger plans further out

Alibaba CEO Eddie Wu mentioned a goal of a model with 5-10 trillion parameters. However, that applies to later generations, Qwen 4.5 and Qwen 5, and not to Qwen 4.

For ordinary users, this means the most powerful Qwen 4 models will likely require payment, while a smaller model may be available to download and use for free.