Alibaba promises new Qwen4 models but has not yet released a date or details

China's Alibaba has been signaling a new generation of its Qwen AI model for several months. The message at the company's Apsara conference in Amsterdam on October 29 was, according to the news site Latenode: "The Qwen4 series is coming soon." No official release date has been announced.

Four variants have been named

According to Yotta Labs, Alibaba has named four Qwen 4 variants: Qwen 4 Max is the flagship, Qwen 4 Plus is a balanced model that can also handle multiple types of input such as images, Qwen 4 Flash is built for fast responses, and Qwen 4 27B has open weights, according to Yotta Labs. Open weights mean the model itself can be downloaded and used by others.

It is unclear when the names were announced. Yotta Labs says that on September 22, Liu Dayiheng spoke only about training and "very soon" without mentioning tiers, while Latenode ties the four names to Alibaba's event report from Amsterdam on October 29. No specifications, pricing or dates were provided. It has not been settled whether the larger variants will also get downloadable weights, and the sources disagree on whether this has been confirmed for the 27B model.

A preview of the architecture

On August 28, Alibaba released Qwen 3.8-Flash-Next, an experimental preview of the architecture that is expected to underpin Qwen4. According to Yotta Labs, it is a so-called sparse mixture-of-experts model: It has 125 billion parameters in total but uses only 6 billion at a time. It can take images and video as input and handle a context of 262,000 tokens, expandable to 1 million. That roughly corresponds to how much text the model can hold in its head at once.

However, it is only a preview and not the full Qwen 4. There are still no benchmarks, meaning comparative tests, for Qwen 4.

When will it arrive?

According to attendees at the conference in Amsterdam, Alibaba Cloud's chief technology officer Feifei Li said Qwen4 is coming "next week." That would be November 2-8. The statement was relayed by attendees and has not been confirmed by Alibaba in any official channel.

Qwen project lead Liu Dayiheng reportedly said on September 22, according to Yotta Labs, that the model was in training and would be released "very soon." A leak from July about a September launch turned out to be wrong.

Yotta Labs also assesses that an unlabeled model with the codename Kalmia in testing could point to an imminent release. This is unconfirmed.

What is still missing?

Pricing, license terms, benchmarks and API documentation have not been released. According to Yotta Labs, there is no Qwen 4 Coder variant, and Qwen3-Coder remains Alibaba's coding line. Tech Insider also mentions an unconfirmed plan for revenue sharing for large users, based on two anonymous sources.