共有:
Open-Weight Licensing

"Open" no longer
means unconditional.

Alibaba's Qwen was released under Apache 2.0 and became shorthand for "download it, use it for almost anything." Its newest model just switched to a license that requires a paid tier once revenue crosses a line. The same week, Ant Group made the opposite bet. Open weights now mean reading the terms model by model.

AI Navigate Editorial·2026.08.14·7 min read
OPEN WEIGHTS MIT · NO CAP Ling-3.0 Flash $50M / 12 mo Qwen3.8 -Max
01
The Shift

Qwen's Apache 2.0 signal
just changed

The assumption that open weights meant "you barely have to read the license" is quietly breaking down.

Alibaba's Qwen family had long shipped under Apache 2.0, a permissive license with almost no strings on commercial use. Researchers and startups alike could download the weights and drop them straight into a product — that's what made Qwen the poster child for open weights.

But according to reporting from the South China Morning Post, the newest model, Qwen3.8-Max, left Apache 2.0 behind and shipped under a new, restrictive custom license called the "Qwen3.8-Max License". Forkast's analysis calls it "not a gift but a platform play."

Qwen3.8-Max LicenseLing-3.0-Flash (MIT)
New restrictive custom licenseFully permissive MIT license
Commercial license required above $50M in trailing 12-month revenueNo revenue-based restriction
Internal use exempt if not exposed to third partiesNo conditions on any use case
Publisher: AlibabaPublisher: Ant Group (inclusionAI)

"Open" no longer means unconditional.
From here, you check the license model by model, release by release.


02
The Counterexample

The same week, Ant Group
made the opposite call

Almost the same moment Qwen added conditions, Ant Group went the other way entirely.

$50M
Revenue line for Qwen3.8-Max's commercial license
124B
Ling-3.0-Flash total parameters
51B
Activated per generation (MoE)
256K
Context window (tokens)

According to a report from KuCoin, Ant Group's inclusionAI team open-sourced Ling-3.0-Flash under a fully permissive MIT license. It has 124B total parameters, and its MoE (Mixture of Experts) architecture activates just 51B of them per generation. Per a technical writeup from digitalapplied, its context window reaches 256,000 tokens, and it's built around programming, search, and agent workloads.

There is no revenue-based restriction at all. In the same week, one company chose "free, with a ceiling" and the other chose "no ceiling." That's not a coincidence in tone — it reads as the first visible fork in open-weight strategy.

03
How It Works

Where does "commercial" start?

The Qwen3.8-Max License draws its line on two axes: third-party exposure and revenue scale.

Use the model Exposed to a third party? No Internal use is exempt Yes Trailing 12-mo revenue over $50M? No Stays free to use Yes Commercial license required
FIG. The Qwen3.8-Max License decision flow — third-party exposure and $50M in trailing 12-month revenue are the forks
01

Internal use is exempt

As long as the software, its outputs, or the underlying model's capabilities aren't made available to third parties, internal use isn't bound by the license's restrictions.

02

Check for "model as a service"

If you (or an affiliate) run a model-as-a-service or "AI work assistant" business exposed externally, that entity's trailing 12-month aggregate revenue becomes the test.

03

Cross $50M and a commercial license kicks in

Once aggregate revenue over any consecutive 12-month period exceeds $50 million, a commercial license agreement with Alibaba becomes required.

04
Who It Affects

What this means for you, by role

Engineer

If you've been fine-tuning Qwen assuming Apache 2.0 terms, you now need to check the license per version starting with Qwen3.8-Max. Keep an MIT-licensed alternative like Ling-3.0-Flash in your evaluation set so a swap doesn't stall a launch.

Business

$50 million a year isn't a distant number. If your model-as-a-service or "AI work assistant" business built on Qwen3.8-Max is approaching that line, start the commercial licensing conversation with Alibaba early — cost models built on "it's free" assumptions need a second look.

Product Manager

Treat licensing as something you version-control on the roadmap. Don't assume the next release carries the same Apache 2.0 terms as the last one just because it used to. Keep an unconditionally-licensed backup model on the shortlist at all times.


05
Risk & Outlook

This isn't purely alarming

Still, this isn't cause for panic. The $50 million threshold is out of reach for most individual developers and small teams for now, and internal use is explicitly exempt. As Forkast puts it, this isn't so much "the end of free distribution" as a mechanism for Alibaba to capture revenue from commercial users who succeed with the model.

Real uncertainty remains. How strictly the "AI work assistant" language will be enforced, and how affiliate revenue gets aggregated, are questions that will only be settled by how the license is actually applied. And Ling-3.0-Flash's MIT license is no guarantee Ant Group won't change terms on a future release — what this split really demonstrates is that the old assumption of "open weight equals permanently unconditional" no longer holds.

Three concrete actions follow. First, audit which open-weight models you rely on and their exact license, version by version. Second, if commercial revenue is approaching the threshold, reach out to Alibaba's commercial licensing team early. Third, keep unconditionally-licensed alternatives like Ling-3.0-Flash evaluated and ready as a fallback.