On 3 August 2026 the Qwen team announced Qwen3.8-Max, a 2.4 trillion parameter mixture-of-experts model with roughly 95 billion parameters active per token, and said the open weights would follow within a week. That is the first time a Max-class Qwen model had been promised as an open release, and promised releases are exactly the sort of thing the watchlist exists to track. Announcements are free; uploads are not.

The weights are up. The repository Qwen/Qwen3.8-2.4T-A95B on Hugging Face was created on 8 August 2026 and last modified on 12 August 2026, with the model card, chat template and licence file all landing across that window. The checkpoint has been publicly downloadable since 12 August 2026, which is the date we resolve on. A companion dense model, Qwen3.8-27B, followed under Apache 2.0.

The watchlist resolves YES, with two qualifications that matter more than the headline.

The first is timing. “Within a week” from 3 August lands on or about 10 August. The weights were live on 12 August. Two days late on a self-imposed deadline is not a broken promise in any meaningful sense, and we do not think it should be scored as one — but the watchlist records the date, not the sentiment, and the date slipped.

The second qualification is substantive, and it is the reason we mark this partial rather than clean. What was released is not the model you call through the API. The model card is explicit: Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, including vision input, non-thinking support, a one million token default context and built-in tools. The open checkpoint is text only. It cannot take images. It requires thinking mode and cannot run without reasoning. Its native context is 262,144 tokens, extensible to about 1,010,000 rather than default. The licence is not Apache 2.0 but a custom qwen3.8-max licence carried in the repository, unlike the 27B model beside it.

So the honest reading is that Alibaba open-weighted the base of a Max-class model rather than the Max product. That is still a first, and a substantial one — a 2.4 trillion parameter checkpoint of this class has not previously been released by anyone under any licence. Teams with the hardware to serve it now have a text model in the frontier tier that they can host, audit and fine-tune themselves, which is a different proposition from renting one. Anyone reading the announcement as “the API model, but free” will be surprised by the missing vision encoder and the reduced default context.

The custom licence deserves a read before anyone builds on it. A bespoke licence attached to a flagship release is a deliberate choice, and it is a different risk profile from the Apache 2.0 terms on the 27B model in the same family. We have not yet finished a clause-by-clause read and will not characterise its restrictions until we have.

One watchlist item stays open behind this one: the self-reported Qwen3.8-Max benchmark figures have not been independently reproduced. Open weights make that reproduction possible for the first time, but only for the text-only checkpoint, which is not the configuration the published scores were measured on. Reproduction against the open weights will therefore test something adjacent to the claim rather than the claim itself. We will re-check that item at the end of September 2026.

Sources