Qwen 3.8 Max promised open weights. Where are they?

Alibaba released Qwen 3.8 Max on 3 August 2026. It is a 2.4-trillion-parameter model with a million-token context window, and the company says it sits second only to Fable 5 (Yotta Labs on release and access).

It also came with a promise. The weights would go public "next week" on Hugging Face and ModelScope. That week has arrived, and so far the shelves are bare.

Key Takeaways

• Qwen 3.8 Max launched on 3 August 2026 as an API-only model at $2 per million input tokens and $6 per million output tokens, with weights promised the following week.
• A check of Hugging Face's public model API on 7 August 2026 found no official Qwen3.8 repository, and Alibaba has still not named a license.
• Even when the weights land, the 2.4T flagship is far beyond consumer hardware, so the unreleased 27B model is the one most developers should watch.

What Alibaba actually shipped

The launch was an API launch. You can call Qwen 3.8 Max today, and you can read its price list, but you cannot download it.

The published numbers are specific. Input runs $2.00 per million tokens, output $6.00, and cached input $0.25, and the context window is one million tokens with a maximum output of 128,000 (Developers Digest on the launch).

Open weights are model files any developer can download and run on their own hardware. That is a different thing from an open API, and it is the part Alibaba has not delivered yet.

Detail	Status as of 7 August 2026
API access	Live since 3 August
Parameters	2.4 trillion, reported as a mixture-of-experts design
Context window	1 million tokens, 128k maximum output
Price	$2 in / $6 out / $0.25 cached, per million tokens
Open weights	Promised "next week" from 3 August; not published
License	Not named

The weights are not on Hugging Face

We checked this directly rather than taking it on trust. On 7 August 2026 we queried Hugging Face's public model API for repositories owned by the Qwen organisation, sorted by most recently modified. The newest public entry was from 22 July 2026, and no Qwen3.8 repository appeared at all.

Four third-party repositories do exist, such as an FP8 build that names Qwen/Qwen3.8-27B as its base model. They were created on 5 and 6 August, and every one of them holds exactly two files: a README and a .gitattributes. There are no weights in any of them, because there is nothing yet to quantise.

That detail is worth sitting with. People created placeholders for a model that does not exist publicly, tagged one of them Apache 2.0, and waited. The license tag is that uploader's assumption. Alibaba has named no license at all.

The benchmark table is thinner than the claim

"Second only to Fable 5" is a big statement. The evidence Alibaba published alongside it is narrower.

The vendor-reported figures cover PaperBench at 93.0, CoWorkBench at 74.8, and WideSearch at 81.9. Those are the company's own numbers, released without an independent audit (Latent Space's launch roundup).

Third-party results have started arriving since, and they are respectable rather than record-breaking. SWE-bench is a test that asks a model to resolve real software issues from open repositories, and Qwen 3.8 Max reports 87.3% on it. Public leaderboards also place it fourth overall on Frontend Code Arena at 1,668 Elo and at 66.1 on the Vals Index (Techsy's launch analysis).

None of that makes the model weak. It does mean the ranking claim is currently the vendor's, and the public scoreboards put it near the top rather than at it.

Why the 27B matters more than the flagship

Here is the part that gets lost in the parameter count. Almost nobody reading this will ever run a 2.4-trillion-parameter model locally, whatever license eventually ships with it.

A model of that size needs server-class accelerators and a great deal of memory, and memory is exactly what the market is short of right now. The same ram shortage pushing up phone and PC prices sets the floor for anyone trying to self-host frontier-scale weights.

So the meaningful release is the smaller one. Alibaba announced a Qwen3.8-27B alongside the flagship, and a 27B model is something a well-equipped workstation can actually hold. Consider which of the two changes anyone's day-to-day: it is not the 2.4T.

Alibaba has published no architecture details, no context length, and no benchmark scores for the 27B (ofox.ai on access and open weights). It is the more useful model and the less documented one.

Open weights without a license is not open

A license is not paperwork. It determines whether a company can put the model in a product, whether a startup can fine-tune it, and whether any of it is safe to depend on.

ModelScope is Alibaba's own model hub, and it is the second destination named in the promise. Publishing to both hubs without license terms would still leave every commercial user waiting.

This matters more for a Max-class model than for a small one. It would be the first flagship Alibaba has opened, and the terms attached are the whole story for anyone deciding whether to build on it.

Until those terms exist, "open weights" describes an intention rather than a release.

What would change the picture

Three things are worth watching this week.

The first is a repository appearing under the official Qwen organisation, not a third-party mirror. The second is a named license, ideally a standard one rather than a bespoke set of terms. The third is a model card for the 27B carrying the details the flagship announcement skipped.

If all three land, this becomes the most capable set of open weights anyone has published. If the week passes quietly, the more honest description of Qwen 3.8 Max is a competitive commercial API with an open-source announcement attached.

Frequently Asked Questions

Can I download Qwen 3.8 Max right now?

No. As of 7 August 2026 the model is API-only, and no official weights have appeared on Hugging Face or ModelScope.

What license will the weights use?

Alibaba has not said. A third-party placeholder repository tags Apache 2.0, but that is the uploader's guess and carries no authority.

How much does the API cost?

$2.00 per million input tokens, $6.00 per million output tokens, and $0.25 per million cached input tokens, per Alibaba's published price list (Developers Digest pricing detail).

Is Qwen 3.8 Max really the second-best model available?

That claim comes from Alibaba. Independent scoreboards currently place it near the top of open-weight and coding rankings, including 87.3% on SWE-bench, rather than in outright second place overall.

Which model should a developer actually plan around?

The unreleased Qwen3.8-27B. It is the version that fits on hardware most teams own, though Alibaba has published almost no detail about it so far.

The promise is the story

A launch is easy to verify. A promise takes a week to check, and most coverage moves on before the check happens.

Qwen 3.8 Max is a real model with real pricing and a credible showing on public leaderboards. Whether it is also an open model is a question with a deadline, and that deadline is this week.