Alibaba Cloud

Qwen tells you its licence in the model's name, and the tier vocabulary has never once broken the rule. Alibaba Cloud had 53 rows in the OpenRouter catalog, as observed on 31 August 2026, second only to OpenAI's 60. Seventeen carried no Hugging Face id — and every one of those seventeen is named max, plus or flash. The tier-named rows and the closed rows are the same seventeen: not a single max, plus or flash row is downloadable, and not a single row without one of those three words is withheld. Thirty-four of the other thirty-six are named by parameter count as well, but that half of the convention does break — qwen/qwen3-coder carries its 480B only in its display name, and qwen/qwen3-coder-next carries no parameter count in its id, display name or slug at all. Both publish weights, which is what the tier test predicts for a row with no tier word in it. The tier vocabulary Alibaba uses for marketing is doing double duty as the open-source policy, which means you can settle the question about any Qwen row in the snapshot by looking for one of three words.

That makes the pair listed in August worth reading together. qwen/qwen3.8-max went up on 3 August 2026 as the closed flagship of the series. Nine days later Alibaba listed qwen/qwen3.8-2.4t-a95b, described on its own page as "the open-weight variant of Qwen3.8 Max" — 95 billion active parameters out of 2.4 trillion — at $2.00per million tokensopenrouter-models, last checked 2026-09-14 for input, the same figure the closed row carries at not publishedlast known value, as of an unrecorded date — the source no longer lists this rowopenrouter-models, last checked 2026-09-14, with a marginally larger window: 1048576tokensopenrouter-models, last checked 2026-09-14 against not publishedlast known value, as of an unrecorded date — the source no longer lists this rowopenrouter-models, last checked 2026-09-14. Publishing the weights of the flagship cost the hosted flagship nothing on the price sheet, which is not how the trade-off is usually described.

The tier names hold still; what they cost does not. qwen/qwen3.7-flash, listed 27 July 2026, lists at $0.03per million tokensopenrouter-models, last checked 2026-09-14. Its successor qwen/qwen3.8-flash, listed 26 August, lists at $0.15per million tokensopenrouter-models, last checked 2026-09-14 — in the 31 August 2026 snapshot, five times as much thirty days later, on the tier whose whole name is a promise about cost. Both are closed, both take text, images and video, and both advertise the same window, so the increase is not paying for a longer context or a new modality.