Alibaba's Qwen team previewed Qwen3.8-Max-Preview on July 19, 2026, at the World AI Conference in Shanghai, and the headline number is large: 2.4 trillion parameters. Qwen describes it as a sparse Mixture-of-Experts multimodal model that handles text, images, video, and documents with a one million token context window, and its own benchmarking places it second only to Anthropic's Fable 5. The catch, and it is a real one, is that this is a preview. There is no published model card, no license, no full benchmark table, and no disclosed active parameter count. You can try it today through Alibaba's coding surfaces at ten percent of standard pricing, but you cannot yet download it or read the receipts.
The short answer
Alibaba's Qwen team previewed Qwen3.8-Max-Preview on July 19, 2026, at the World AI Conference in Shanghai. It is a 2.4 trillion parameter multimodal model that Qwen ranks second only to Anthropic's Fable 5, with open weights promised soon. You can test it now through Alibaba's Token Plan and its Qoder coding surfaces at ten percent of standard pricing, but there is no published model card, license, or full benchmark table yet.
New flagship models tend to arrive with a benchmark table and a blog post you can pick apart. This one arrived with a teaser. On July 19, 2026, at the World AI Conference in Shanghai, Alibaba's Qwen team previewed Qwen3.8-Max-Preview, a 2.4 trillion parameter model it calls one of the most capable available today. We think the interesting part is not the size claim, which everyone makes, but the shape of the release: a live preview you can use immediately, paired with a promise of open weights and very little published detail to go on.
What Qwen actually showed
Qwen3.8-Max-Preview is a sparse Mixture-of-Experts model, meaning the full 2.4 trillion parameters are the total pool, while only a subset activates for any given token. It is multimodal, accepting text, images, video, and documents, and it carries a one million token context window, enough to hold a large codebase or a long set of documents in a single request. Qwen's own framing puts it second only to Anthropic's Fable 5 among the systems it compared against.
The announcement came straight from the Qwen account, which is worth reading in the team's own words rather than paraphrased.
The preview is not a paper launch. It is live through Alibaba's Token Plan and the Qoder and QoderWork coding surfaces, and Alibaba is charging ten percent of standard pricing during the preview. That pricing choice reads as an invitation to test at scale, which is a sensible way to gather real-world feedback before a full release.
What is missing, and why it matters
Here is the honest gap. There is no model card, no license, no complete benchmark table, and no disclosed active parameter count. For a preview that is not unusual, but it does shape how much weight the claims can carry. The parameter total tells you the model is large. It does not tell you how much compute each query burns, because in a Mixture-of-Experts design that is set by the active parameters per token, and Qwen has not published that figure.
The open-weight promise deserves the same caution. Qwen posted that the model is going open-weight soon, but soon has no date and no license attached. Open weights would let you download the model, run it on your own hardware, fine tune it, and audit its behavior, which is a very different proposition from a hosted API. Until the weights and terms actually ship, the practical stance is to treat Qwen3.8 Max as a preview you can experiment with, not a model you can plan a self-hosted deployment around.
The wider picture
The timing is the story as much as the specs. Qwen's preview landed two days after Moonshot AI released Kimi K3, a 2.8 trillion parameter open-weight model, and both sit in the same weight class with the same sparse Mixture-of-Experts approach. Two frontier-scale models arriving within a week, both promising or delivering downloadable weights, is a signal about where this part of the field is heading.
For developers, that competition is the useful part. When capable models pile up this quickly, the pressure runs toward higher capability and lower prices, and toward more models you can actually run yourself instead of only renting. Qwen3.8 Max is not something to build on today, given how little is published. It is a strong reason to keep the finished release on your radar, and to compare it against the open weights arriving alongside it once the real numbers land.
Sources and further reading
- MarkTechPost: Alibaba previews Qwen3.8-Max, a 2.4 trillion parameter multimodal model
- Bloomberg: Alibaba's Qwen unveils preview of flagship AI model
- The Edge Singapore: Alibaba's Qwen unveils preview of flagship AI model
Frequently asked questions
What exactly did Alibaba announce?
On July 19, 2026, at the World AI Conference in Shanghai, Alibaba's Qwen team unveiled Qwen3.8-Max-Preview, the next flagship in the Qwen family. It is a 2.4 trillion parameter, sparse Mixture-of-Experts multimodal model that accepts text, images, video, and documents, with a one million token context window. Qwen says the model is one of the most capable available today and, by its own benchmarking, second only to Anthropic's Fable 5. Crucially, this is a preview, not a finished release.
Can I use it right now?
Yes, in preview form. Qwen3.8-Max-Preview is live through Alibaba's Token Plan and the Qoder and QoderWork coding surfaces, priced at ten percent of standard rates during the preview window. What you cannot do yet is download open weights, read the license, or check a complete benchmark table, because none of those have been published. So it is usable for testing, but not something you can self-host or fully evaluate today.
How does 2.4 trillion parameters compare to recent releases?
It lands in the same weight class as the other large models from this month. Moonshot AI released Kimi K3 at 2.8 trillion parameters as an open-weight model just two days before Qwen's preview. Both use sparse Mixture-of-Experts designs, where only a fraction of the total parameters activate per token, so the raw parameter count is an upper bound on size rather than a direct measure of the compute each query costs. Qwen has not disclosed how many parameters activate per token, which is the number that actually drives inference cost.
What does open-weight soon mean in practice?
Qwen posted that the model is going open-weight soon, but soon has no date attached, and no license terms are public. Open weights would let teams download the model, run it on their own hardware, fine tune it, and audit its behavior, which is the difference between renting access and owning a deployment. Until the weights and a license actually ship, the responsible read is to treat Qwen3.8 Max as a hosted preview and wait for the real release before planning any self-hosted work around it.
Why should developers outside China care?
Because it is another data point in a fast-moving open-weight race. Two frontier-class models in the 2.4 to 2.8 trillion parameter range appeared within days of each other, both promising or delivering downloadable weights. For anyone choosing a model to build on, that pressure tends to push capability up and prices down, and it widens the set of models you can run yourself rather than only through a paid API. It is worth watching, even if you wait for the finished release before committing.