Qwen 3.8 Max

Alibaba previewed Qwen 3.8 Max on July 19, 2026, at the World AI Conference in Shanghai, and made the hosted model generally available on August 3. It is a mixture-of-experts design with 2.4 trillion total parameters and 95 billion activated per token, drawing on 512 experts of which ten routed and one shared fire for any given token. The published architecture is unusual in the open: 92 layers arranged as repeating blocks of gated DeltaNet and gated attention, each followed by a mixture-of-experts layer.

The Max tier had always been the closed half of Alibaba's catalogue. Smaller Qwen models had shipped with open weights for years, and the Max models were the ones you could only rent. In August the company published the weights as Qwen3.8-2.4T-A95B on Hugging Face under a licence named for the model, making this the first member of the Max class anyone could download.

The downloadable file is not the same object as the hosted service, and the model card says so. The released weights are text only and require thinking mode for every interaction, while the product announced in July was multimodal across images, video and documents. The context window is 262,144 tokens natively, extensible to roughly 1,010,000.

Three weeks earlier Moonshot had published Kimi K3 at 2.8 trillion parameters, and the obvious reading of Qwen 3.8 Max is that a second Chinese laboratory matched the gesture. The more interesting fact is the one about tiers. Kimi K3 was an open model from a company whose models are open. This was a company reaching into the part of its catalogue that had always been rented and putting it on a download page, which is a decision about a business rather than about a release.

It also complicates what an open-weight release means. For most of the period this museum covers, weights published meant the thing itself, and a reader could assume the file and the service were the same model. Here they are not: one is multimodal and hosted, the other is text only and requires a particular mode to run at all. Asking whether a model is open is no longer a question with one answer, and the honest version of the question is which artifact, with which capabilities, under which licence.