Qwen 3.8-Max's Open Weights Have Terms and Conditions Now

On August 3, Alibaba made a promise: the open weights for Qwen3.8-Max — their 2.4-trillion-parameter flagship — would ship "next week." We wrote about it at the time and called it the part that should worry every AI company charging premium API prices.

The weights landed on August 12. Hugging Face, BF16 safetensors, an FP8 variant for people who own serious hardware. The community cheered. The repo racked up a thousand likes in five days.

Almost nobody opened the LICENSE file.

I did. It's not Apache 2.0. It's not MIT. It's a custom "Qwen3.8-Max License" — an MIT-style grant wrapped around two carefully machined clauses that put tolls on exactly the business models that make money off open weights. And when you read it side by side with Kimi K3's license from three weeks earlier, a pattern emerges: the free lunch is over. Not for you, probably. But for someone.

Let me show you what's actually in there, because most of the coverage this week got it wrong in both directions — some called it "open source, no strings," others called it a bait and switch. It's neither. It's something new, and something smarter than both.

What the license actually says

The grant itself reads like MIT. Use, copy, modify, merge, publish, distribute, sublicense, sell, host, fine-tune, create derivative works. Free of charge. So far, so generous.

Then come the conditions.

Condition one: if you use the model in a commercial product with more than 100 million monthly active users, or more than $20 million in monthly revenue, you must "prominently display" the model's name in your user interface. That's an attribution requirement. Weirdly retro — it's the kind of clause you last saw in 2010s open source, except it only kicks in once you're enormous.

Condition two is the real one. If you run a "Model as a Service" business or an "AI Work Assistant" business, and your aggregate revenue (including affiliates) exceeds $50 million over any trailing twelve months, you need a separate license from Qwen before using the model commercially at all. There's a carve-out for internal use, as long as you don't expose the model, its outputs, or its capabilities to third parties.

Two things in that clause deserve your attention.

First, the definitions are surgical. "Model as a Service" means giving third parties access to inference or fine-tuning with "meaningful control over the inputs, parameters, or training data." That covers Together, Fireworks, and anyone else hosting the weights for customers. But it explicitly excludes "the mere relaying of requests to models hosted by other third parties." Read that again: if you're a router that forwards requests to Alibaba's API, you owe nothing. If you download the weights and serve them yourself, you owe a license once you're big enough. Someone drew a line straight through the middle of the inference market, and the line is drawn by who hosts the GPUs.

Second — and this is the part nobody expected — the license defines an "AI Work Assistant" as "an independent AI-powered product primarily designed for AI-assisted coding or office productivity," and then, in the actual license text, names Qoder and QwenWork as examples. Qoder is Alibaba's own coding agent. QwenWork is Alibaba's own office assistant. They wrote the clause by listing their competitors' product category and their own products in the same breath. If you're building a Cursor-style coding tool above $50 million in revenue and you want to serve Qwen3.8-Max to your users, Alibaba would like a word. There's a contact email at the bottom of the license: model-business@notice.qwencloud.com. Business, notice. The subdomain is doing honest work.

The week the split happened

Here's what makes this a story instead of a footnote. Look at what shipped in a single four-day window:

• August 11: GLM-5 from Zhipu. MIT license. No conditions.
• August 12: Qwen3.8-Max. Custom license. Revenue riders.
• August 13: DeepSeek V4-Pro. MIT license. No conditions.
• August 14: Qwen3.8-27B. Apache 2.0. No conditions.

Two labs shipped frontier-scale weights clean. One lab shipped clean and dirty in the same week — Apache on the 27B, custom terms on the flagship. That's not inconsistency. That's a strategy.

Qwen is running the enterprise software playbook: the free tier is genuinely free, the commercial tier has commercial terms. The 27B exists to commoditize everyone else's mid-market and vacuum up adoption — it hit 267,000 downloads in three days and Simon Willison, who tested it the day it dropped, called it excellent while documenting that its default reasoning effort burned 22,000 thinking tokens drawing a pelican on a bicycle. The Max exists to sit in every "world's best open model" leaderboard while quietly refusing to be the thing that makes its users rich without a check.

And the download numbers tell you exactly what people do when given the choice: the 2.4T Max has been pulled 7,900 times. The 27B has been pulled 267,000 times. People download what they can run, and they run what's unencumbered.

Kimi got there first, and it matters

Moonshot AI shipped Kimi K3 on July 27 — 2.8 trillion parameters, the largest open weights ever released — under a custom "Kimi K3 License" with the same skeleton: MIT-style grant, then a clause requiring a separate agreement for Model as a Service businesses above $20 million in trailing-twelve-month revenue.

Note the differences, because they're the interesting part. Kimi's threshold is $20 million. Qwen's is $50 million. Kimi has no attribution clause and no "AI Work Assistant" category. Qwen added both, raised the toll gate, and aimed one of the gates at a specific product category. The structure is converging while the terms iterate. This is version two of a template. There will be a version three.

Two of the three biggest Chinese open-weight labs have now shipped custom revenue-share licenses on their flagships in the same month. DeepSeek and Zhipu are still shipping MIT. Either they're principled, or they're early. The optimist's case: DeepSeek has made "actually open" part of its brand and is winning downloads on it — Kimi K3 has 2.1 million. The cynic's case: every one of these labs is burning compute subsidizing the world, and the first time a CFO runs the numbers, $50 million thresholds start looking like mercy rather than a toll.

The weights aren't even the whole model

One more detail from the repo that got less attention than the license: the downloadable Qwen3.8-Max is text-only, with thinking you can dial down but cannot switch off — the chat template literally raises an exception if you try to disable it. The model's architecture string is qwen3_5_moe_text. The vision input and the 1-million-token context window that made the launch notable? Those live in the hosted API. Same name on the box, different artifact inside.

So the full picture of what Alibaba shipped: the reasoning core, under a license with tolls, missing the multimodal parts that make the product the product. If what you wanted was "download the frontier and never pay anyone again," you got maybe 70% of that, with an asterisk.

To be fair — and this is where the coverage screaming "bait and switch" loses me — 70% of a frontier model, free, for everyone under $50 million, is still an absurd deal by any historical standard. Two years ago this artifact would have been the single most valuable file on the internet. Now it's a Tuesday release with a terms-of-service dispute. That's progress, wearing a suit.

And then Stripe bought the toll road

Four days after the Qwen release, on August 16, Stripe closed a deal to buy OpenRouter for over $7 billion. OpenRouter is the biggest model router in the market — one API in front of hundreds of models, the plumbing half the AI tooling ecosystem runs on.

Now reread Qwen's license with that news in the window. The clause defines Model as a Service. It carves out "mere relaying of requests." It builds a category system for the entire inference stack: relay and you pay nothing, host and you pay above $50 million, embed in a big product and you display the badge, build a coding assistant and you negotiate.

Someone at Alibaba wrote a license that maps the AI business landscape with the precision of a payments network's fee schedule — and four days later, a payments network paid $7 billion for a company that sits precisely on the free lane of that map. OpenRouter the router owes Alibaba nothing. OpenRouter's hosted competitors owe Alibaba a conversation. I don't know how much of this is coordinated strategy and how much is convergent opportunism, but the license and the acquisition are describing the same economy, and neither was written by people guessing.

What this means for you

Concretely, practically, in order of how much you should care:

If you're an individual or a small team: nothing changes. You can run Qwen3.8-Max, build products on it, sell those products, fine-tune it, ship it to customers. The thresholds are 100 million monthly users, $20 million a month, $50 million a year. If those numbers describe you, you have lawyers for this. For everyone else the license is, functionally, more permissive than GPL software you already ship.

If you're picking models for a product: license is now an eval criterion. It belongs next to context window and price per million tokens. A model's benchmark score no longer tells you what it costs your business model — the LICENSE file does. That's genuinely new, and it means model selection decisions made today might have contract implications in two years when you're bigger and the model is load-bearing.

If you're building an inference business or a coding assistant: you already knew. This is just the second data point confirming the first.

The bigger shift is philosophical, and it happened quietly. "Open weights" used to be a binary. Now it's a tier. Free-for-everyone, free-with-attribution, free-until-you're-big. The word "open" is doing less work every quarter, and the honest vocabulary is becoming something like "weights with means-tested pricing." Alibaba would hate that phrase. It's also exactly what the license says.

My prediction, for the record: Meta ships a similar structure within a year — they've watched two Chinese labs test it in production without a community revolt, and the backlash has been remarkably muted because the thresholds are set above almost everyone's head. DeepSeek holds at MIT long enough for it to become a moat, then quietly revisits. And the next Max-class Qwen release doesn't make headlines for its license anymore, because by then it's just how flagships ship.

The 27B stays Apache. Someone has to keep the funnel full.

---

If this stuff matters to you because you actually run these models rather than read about them — CopperRiver is a desktop AI assistant that runs open-source models locally and automates real work: browsing, terminal, files, the unglamorous stuff. The models it uses are ones you can still download without a business development conversation. Take a look.