The Qwen3 Max family

Which version belongs in deployment

If your input is text only and your workload ships in batches, hold on 3.7 Max. If you feed images or video, require structured outputs, or need an answer longer than 65,000 tokens, 3.8 Max belongs in deployment. This is the only family in this reference where the answer to which version is not a single identifier.

The data is extracted from each model dedicated documentation: both carry a 1,000,000-token context window and both cap input at 991,808 tokens, so this parameter creates no distinction. The output ceiling is 65,536 tokens on 3.7 and 131,072 tokens on 3.8, exactly double. The chain-of-thought ceiling is 262,144 tokens on both.

The capability table is where the real distinction lives. 3.7 Max accepts text input only, and its documentation describes it as a pure text interface; 3.8 Max accepts image, text, and video input. Structured outputs are inactive on 3.7 and active on 3.8. And Batch Inference runs the exact opposite: active on 3.7, inactive on 3.8.

A capability removed from the architecture during the upgrade

This last parameter is worth examining separately. The 3.7 Max documentation does not simply log Batch Inference as active; it carries an independent rate for it: batch file input priced at $0.825 and batch output at $2.475, half the standard rate. In the 3.8 Max documentation, Batch Inference is logged as inactive across all six regions, and no batch rate row exists at all.

For any architecture processing several thousand documents overnight in batches, this means migrating to 3.8 doubles the unit cost, with no figure changing in the headline rate column. This class of regression does not surface in standard comparison lists, because those lists check only the rate column, not the capability column.

Rate depends on deployment region, and a temporary discount sits in the middle of it

The 3.8 Max documentation prices six deployment regions independently. Beijing, Frankfurt, Virginia, Tokyo, and Hong Kong all carry $1.65 input and $4.951 output. Singapore, logged as the international region, carries $2 and $6. The figure recorded in our tables is that international row.

On the central pricing page, 3.7 Max shows a list rate of $2.5 and $7.5 in the international row, with a 50 percent limited-time discount label beside it. So deployment today may cost half that rate, while budgeting six months out may not carry the discount at all. We always use the list rate as the basis for calculation, since a temporary discount is not a figure a long-term contract can be built on.

How this model was located, and why that process is itself part of the content

Until a few days ago, the rate and context-window columns for this model were logged as unverified in our table, and the reason was documented explicitly: the English Model Studio catalogue and the central pricing page, both last updated 15 July 2026, did not log the identifier qwen3.8-max, while the Arena leaderboard row carried exactly that identifier. Assuming 3.8 was simply 3.7 would have been the easy assumption, but it was a guess, and once the two documents are placed side by side, that guess is confirmed incorrect.

What further investigation found: a dedicated page exists for this model, updated 3 August 2026. So Alibaba index pages had fallen behind their own models, rather than the model not existing. One detail makes this hard to locate: the URL slug is written with a hyphen while the model identifier uses a dot.

The operational conclusion for any analysis checking the specifications of a Chinese model: the vendor index is not the final source. The model dedicated page address must be checked, even when it is absent from the index.