Alibaba's most 'open' model yet comes with a revenue-share clause. Read the $50 million line.

Alibaba released Qwen3.8-Max this week and called it the company's most capable model yet, the latest and largest entry in a family it has spent two years positioning as the open alternative to the closed American labs. The technical claims are substantial: 2.4 trillion total parameters, roughly 95 billion active at any one moment, a context window of a million tokens. The framing around it is the familiar one — a commitment to open weights, a gift to the developer community, a model anyone can download and run. Read the licence that ships with it, and the framing acquires an asterisk.

For the flagship, Alibaba is introducing something it has never attached to an open-weight Qwen before: a revenue-share requirement. Large commercial users — companies running the model as a paid service — will owe Alibaba a cut. The exact percentage is reported to be still under negotiation. The threshold that triggers it is already specific, and the threshold is where the strategy is written.

What 'open' just conceded

Until now, the open Qwen models shipped under Apache 2.0 — one of the most permissive licences in software. Apache 2.0 asks almost nothing: use it commercially, modify it, build a business on it, keep the profits, no payment and no permission required. Alibaba monetised the arrangement the way most open-model providers do, by charging for the convenience of running the model on Alibaba Cloud's own API while letting anyone who preferred to host it themselves do so for free. The generosity was real, and it was also a customer-acquisition strategy. Free weights build a developer base; the developer base rents cloud.

Qwen3.8-Max keeps the shape of that arrangement and adds a trapdoor. The flagship no longer ships under Apache 2.0. It ships under a custom licence with a commercial condition, and that condition is the news. 'Open' has not been withdrawn — you can still download the weights, still run them, still build on them. What has changed is that past a certain size, building on them now costs. The word survived. The terms under it did not.

The $50 million tell

The revenue-share obligation is reported to apply to users — or their affiliates — running a 'model as a service' or 'AI work assistant' business whose aggregate revenue exceeds US$50 million over any consecutive twelve-month period. Read that clause slowly, because it is doing precise work. It does not touch researchers. It does not touch the startup running Qwen on a rented GPU. It does not touch the enterprise using it internally. It touches, specifically, companies that have turned Qwen into a product large enough to matter — and it touches the ones reselling it as a service most of all.

That is not a clumsy net that happens to catch big fish. It is a net cut to the exact shape of one kind of fish. A $50 million services business built on Qwen is, almost by definition, a competitor to Alibaba Cloud — someone hosting and selling access to the same model Alibaba would prefer you buy from Alibaba. The revenue share is not really a charge for using Qwen. It is a charge for using Qwen to compete with the part of Alibaba that gives Qwen away.

The revenue share is not a charge for using Qwen. It is a charge for using Qwen to compete with the part of Alibaba that gives Qwen away.

Two licences, one strategy

Alibaba did not put every model behind the new terms, and the split is deliberate. The smaller Qwen3.8 weights, reporting indicates, continue to ship under the permissive licence; only the flagship carries the commercial condition. That is the tell that this is a strategy rather than a retreat. The goodwill — the researcher downloads, the leaderboard wins, the reputation as the open lab — lives on the small and mid-size models, where nobody is building a $50 million business anyway and the free licence costs Alibaba nothing. The monetisation lives on the flagship, where the users large enough to pay are the only ones affected. You keep the halo and meter the frontier. It is a clean piece of segmentation.

It also quietly discontinues something, and the discontinuation is the honest document. For two years the promise attached to Qwen was Apache 2.0 on the good models — open in the strong sense, no strings, including on the best thing Alibaba made. That promise is the thing that just ended, on the flagship, without being announced as an ending. The launch describes a more capable model. The licence describes a more expensive one, for anyone it is designed to reach.

Meta already wrote this playbook

None of this is unprecedented, which is part of the point. Meta's Llama models have carried a similar gate for years — open to nearly everyone, with a clause that flips for companies above a set scale of monthly users, aimed squarely at the handful of rivals large enough to turn Llama against Meta. Alibaba has now adopted the same structure with a revenue threshold in place of a user count. The convergence matters: the two most prominent 'open' model families in the world have arrived at the same design, in which open means open until you are big enough to be a threat, and then it means negotiate.

That is worth stating plainly, because the industry has spent two years treating 'open weights' as a single, stable category — the counterweight to the closed labs. It is not a single category. It is a spectrum of licences, and the most valuable models on it are drifting toward the restricted end at exactly the point where the licence starts to matter commercially. The weights are open. The business model attached to them is not.

The strategy the launch won't state

Stripped of the framing, what Alibaba is running is a funnel with a hedge at the end of it. The free tier — permissive small models, downloadable flagship weights — does what free tiers do: it acquires developers, wins benchmarks, and makes Qwen the default anyone reaches for. The revenue share at the top does the hedging. If a customer runs Qwen on Alibaba Cloud, Alibaba is paid for the compute. If a customer grows large enough to run it themselves and sell it on, Alibaba is paid the revenue share. Either way, past a certain scale, Alibaba collects. The company has arranged things so that Qwen's success accrues to Alibaba whether or not Qwen runs on Alibaba's infrastructure.

That is a genuinely smart design, and it is not the design the word 'open' describes. It is closer to a platform toll than to a gift — free at the bottom to build the base, metered at the top to capture the winners, with the cloud business hedged on both sides. Read that way, the $50 million threshold is not a limit placed on an open model. It is the price of admission to the only tier where Alibaba was ever going to collect.

The announcement is about capability, and the capability is real — Qwen3.8-Max is a serious model, and by several measures the strongest open-weight system anyone has shipped. But the document that tells you what Alibaba is actually doing is not the benchmark table. It is the licence, and the number in the licence is $50 million. The company would rather you read the first document. The second is the one under oath.