Qwen3.8-Max Is Text-Only — Nearly Everyone Reported Otherwise

Alibaba said at launch on 3 August that it would publish Qwen3.8-Max's weights. On 12 August it did — Qwen/Qwen3.8-2.4T-A95B and an FP8 checkpoint, followed by a separate 27B companion model.

At 2.4 trillion total parameters with 95 billion active across 512 experts, this is now the largest genuinely downloadable model on our roster, ahead of K-EXAONE 2.0 at 750B.

Read the licence before you build on it

The weights are not Apache 2.0. They are published under a licence named qwen3.8-max, so the blanket commercial permission that comes with K-EXAONE 2.0 or Command A+ does not automatically apply here. Check the terms against your use case.

Two corrections to our own entry — and to nearly everyone else’s

Alibaba’s model card contradicts the launch coverage we originally seeded from, and we have updated the roster entry accordingly:

Context window. Most coverage reports a flat 1M tokens. The model card documents 262,144 tokens natively, extensible to about 1.01M.

Vision input — it is text-only. Bloomberg, SCMP and MarkTechPost all described Qwen3.8-Max as natively multimodal, accepting text, images and video, and the API aggregators repeat it. An independent verification pass we ran came back insisting the same thing, citing a Hugging Face card that says "Causal Language Model with Vision Encoder".

That card belongs to a different model. Side by side:

Alibaba’s card for the Max states: "Multimodal inputs are not supported." Alibaba shipped a multimodal Qwen3.8 at 27B in the same family at the same time, and the coverage collapsed the two models into one. That single conflation explains the licence error too — the Apache 2.0 everyone cites belongs to the 27B, or to the GitHub code repository, neither of which governs the 2.4T weights.

Our roster entry records it as text-only with the model card as the source. In this family, Qwen3.8-27B is the model that takes image input.

The card also states that thinking mode is required for all interactions, not optional. Responses always include reasoning before the final output, and you cannot opt out of paying for it.