The Qwen3.8-Max License Bills Your Company, Not Your Inference

Alibaba's first downloadable Max-tier model ships under a license whose two conditions are measured on your total revenue and your product category, not on how you run the weights. Here is how to check whether it lands on you.

The Qwen3.8-Max License is 3.39 kilobytes long and its grant clause is more permissive than most people expect. You may use, copy, modify, merge, publish, distribute, sublicense, sell, deploy, host, fine-tune, and create derivative works from the weights, free of charge. That is MIT phrasing with extra verbs bolted on for the parts MIT never had to name.

Then come two conditions, and neither one is about inference. One is about how many people use your product. The other is about how much money your company makes and what category a lawyer would put your product in. Nothing in either condition asks where the weights run, what hardware serves them, how many tokens you push through, or whether you quantized them down to something that fits on a workstation.

That is the change worth paying attention to. Every argument about open weights for the last two years has been an argument about serving: can you self-host, can you fine-tune, can you run it air-gapped. Alibaba answered yes to all of that and then wrote its conditions against a completely different axis.

The unit of measurement moved

Meta's Llama licenses put a threshold at 700 million monthly active users, and that number measured product scale. If you were big enough to matter, you had to ask.

Qwen's second condition measures something else entirely. Read it carefully:

If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, and the aggregate revenue of the licensee and its affiliates exceeds US$50,000,000 (or the equivalent amount in any other currencies) during any consecutive twelve (12) months, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose.

The threshold is aggregate revenue of the licensee and its affiliates. Not revenue attributable to Qwen. Not revenue from the feature the model powers. Not revenue from AI at all. Your whole company's top line, across twelve consecutive months, in any window.

Work through what that does. A developer-tools company at $60 million a year that uses Qwen3.8 for one autocomplete feature earning nothing directly needs to go get a separate license first. A pure Qwen reseller at $10 million a year does not. The company using the model least is the one that has to ask permission, because the trigger was never about the model.

My position: this is the more consequential open-weights license of 2026, and not because it restricts more than its peers. It restricts differently. It attaches the obligation to your P&L and your product category rather than to your deployment, which means your infrastructure team cannot answer the compliance question at all. Your finance team can.

What the two conditions actually say

Condition one is an attribution rider with a scale trigger. If the software or any derivative is used for a commercial product or service with more than 100,000,000 monthly active users or more than US$20,000,000 in monthly revenue, the model name must be "prominently displayed on the user interface" of that product. Note the or. Either threshold fires it. Note also that this is a UI requirement, not a notice file buried in a legal page. Whatever your product's screen is, the model name goes on it.

Condition two is the separate-license requirement, quoted above. It has one carve-out, and the carve-out is where most teams will actually land:

The foregoing requirement shall not apply to the licensee's internal Use of the Software, provided that such Use does not make the Software, its outputs, or its underlying model capabilities available to any third party.

Internal use is clear of it. But "outputs" and "underlying model capabilities" are broad words. A support system that drafts replies your customers read is arguably exposing outputs to third parties. An internal analytics tool nobody outside the company touches is not.

The definitions section is where the license does its sharpest work, and it is written with more precision than these documents usually get.

"Model as a Service" means giving a third party access to language model inference or fine-tuning, by API or hosted endpoint, in a way that lets that third party "exercise meaningful control over the inputs, parameters, or training data." Then the exclusion: this "does not include the mere relaying of requests to models hosted by other third parties." So if you route your users to Together or Fireworks, you are a relay and not MaaS. If you stand the weights up yourself and hand out API keys, you are MaaS.

"AI Work Assistant" means an independent AI-powered product primarily designed for AI-assisted coding or office productivity. The license names two examples, and they are Alibaba's own: Qoder and QwenWork. It then excludes three things: single-purpose tools like an AI translator, assistants built for a domain other than coding or office work (the license names Taobao AI Shopping Assistant and AMap AI Chat), and an AI assistant that is a feature inside a product whose primary purpose is something else.

Read that exclusion list backwards and you get a map. Alibaba drew a fence around exactly one category, coding agents and office copilots, and that category is the one it competes in. Everything else it named as outside. This is not a general restriction that happened to catch coding tools. It is a restriction aimed at coding tools that carefully lets everything else through.

Correcting the coverage

A lot of the writing about this release calls it a revenue-share license, and several outlets have compared it to reported percentage splits on other Chinese open-weight models.

I read the full license text. There is no percentage in it. There is no revenue-share mechanism, no royalty rate, no formula, no schedule. There is a requirement to "obtain a separate license from Qwen" and an email address, model-business@notice.qwencloud.com. Whatever Alibaba charges is a negotiation that happens off the page.

Reporting that Alibaba plans revenue sharing may well be accurate as reporting. It is not in the document you agree to when you download the weights, and if you are making a decision, the document is what binds you. Do not price a deal against a percentage that exists in a news article.

One more thing worth knowing before you build on it. The downloadable checkpoint is Qwen3.8-2.4T-A95B, roughly 2.4 trillion total parameters with about 95 billion active, and it shipped text-only. The vision capability and the million-token context window that made the hosted Max product notable are not in it. The license is titled Qwen3.8-Max and the Hugging Face metadata tag reads qwen3.8-max, but the artifact under that name is smaller than the product the name refers to. That gap drew loud complaints on the model's own discussion page, and the complaints are fair.

Put this into practice

This is a twenty-minute review, and you can do the first four steps yourself before anyone bills you an hour.

1. Read the license file, not a summary. It is 3.39 kB. It fits on one screen. Open the LICENSE on Hugging Face and read all of it, including the definitions. Most of the confusion in circulation comes from people reading a headline about it.

2. Write down three answers. Do you conduct a Model as a Service business by the license's definition (you host inference or fine-tuning and third parties control inputs, parameters, or training data)? Do you conduct an AI Work Assistant business (an independent product primarily for AI-assisted coding or office productivity)? What was your company's and your affiliates' aggregate revenue across the last twelve consecutive months? Written answers, in a document, dated. If a lawyer ever picks this up, that document is the starting point.

3. If either category fits and you are anywhere near $50 million, stop before you ship. The license says obtain the separate license before Using the software for any commercial purpose. That word ordering matters. Retroactive is not a path the text offers.

4. If you might cross 100M MAU or $20M monthly revenue, put the model name in the UI now. It costs you one string in a footer or an about screen. Adding it later, after you cross, is a compliance conversation. Adding it today is a pull request.

5. Record the commit. The LICENSE landed in commit 207bd68 ("Update README and add LICENSE"). Note that hash alongside your review. Model repositories get updated, and you want to be able to show which text you agreed to.

If your answer to both category questions is a clean no, or you are comfortably under the thresholds, you are running under something close to MIT and you can get on with it. That is genuinely most teams.

Honest limitations

I am not a lawyer and this is not legal advice. What I have done is read the operative text closely and report what it says. If money depends on the answer, pay someone whose job this is.

Three parts of the license are underspecified in ways that will matter to someone.

"Affiliates" is never defined. For a company with a parent, a holding structure, or private-equity ownership across a portfolio, that word could swing the aggregate-revenue calculation by an order of magnitude. There is no ownership-percentage test in the document.

"Prominently displayed on the user interface" has no standard. No size, no placement, no persistence requirement. A footer probably counts. A tooltip probably does not. Nobody can tell you where the line is until Alibaba enforces it somewhere.

And the price of the separate license is unpublished. You cannot budget for it. You can only email and ask, which puts every company in the AI Work Assistant category into a bilateral negotiation with a competitor who knows what your revenue is because you just told them.

Worth saying plainly: this does not meet the Open Source Initiative's definition. It carries a field-of-use restriction and a scale-based obligation, and both are disqualifying. "Open weights" is the accurate term. "Open source" is not, and the distinction is doing real work here.

What to watch

The interesting question is not whether Qwen3.8's license is fair. It is whether "AI Work Assistant" survives as a licensed category.

Alibaba wrote a definition precise enough to name its own products as the examples and its own non-competing products as the exclusions. If that phrasing gets copied, and license text does get copied, then the next generation of open-weight releases will arrive with a category fence around whichever market the releasing lab is currently fighting for. Weights stay free. The one use that would hurt the publisher gets a toll booth.

If you build coding agents on open weights, go read the license on every model in your stack this week and check what the conditions are measured against. Not what they forbid. What they measure. If any of them measure your company rather than your deployment, I would like to know which ones, because I have only found this one so far.

Sources: the Qwen3.8-Max License on Hugging Face (full text, commit 207bd68), the Qwen/Qwen3.8-2.4T-A95B model repository, and the model's community discussion on the text-only release. Secondary coverage of the release from Forkast and llm-stats; reported revenue-share plans are theirs, not the license's.