Qwen's Licence Has Two Conditions. Only One Has a Threshold.

Most people read a model licence by scanning for a number. Qwen Community License 1.0, which covers Qwen3.8-Flash-Next, has one prominent number: 100,000,000 monthly active users. Read only that and you conclude the licence is permissive for anyone smaller, which describes almost everyone.

That number belongs to the condition that will never apply to you. The condition that might apply carries no threshold at all.

What each condition says

Condition 2 reads: if the licensee or any of its affiliates conducts a model-as-a-service or AI work assistant business, the licensee shall obtain a separate licence from Qwen before using the software or its derivative works for any commercial purpose. No revenue floor qualifies it. No user count qualifies it.

The definitions do the work

Two defined terms decide whether Condition 2 reaches you. Qwen defines model as a service as providing third-party access to language model inference or fine-tuning. It defines AI work assistant as an independent AI product for coding or office productivity.

The second definition is broad. A coding assistant is an AI work assistant. A document drafting tool is an AI work assistant. A meeting summariser sold as its own product may be too. Teams building those things are the ones most attracted to a cheap, capable open-weight model, which is the point of writing the clause that way.

Internal use stays clear. Run Flash-Next inside a product that does something else, serving your own users, and neither condition binds you. The line falls between using the model and selling the model.

Qwen split the family across licences

The permissive licence sits on the smaller model. The conditional one sits on the model with the better benchmark scores and the architecture preview attached. Alibaba earlier released Qwen models under Apache 2.0 across the board, so this split is a change in approach rather than a continuation.

For a team choosing between the two, the licence now belongs in the comparison next to the benchmark scores. We put both side by side in Qwen3.8-27B vs Flash-Next, where the dense model is within 0.8 points on SWE-bench Pro and carries no commercial condition.

What to do about it

• Work out which side of the line your product sits on before you benchmark anything. Selling access to the model is the trigger, not model size or revenue.
• If you build coding or office productivity tools as standalone products, read Condition 2 in full and price in a conversation with Qwen.
• If the model runs inside a product that does something else for your own users, neither condition applies.
• Where the licence is the blocker rather than the capability, check whether the Apache 2.0 tier clears your bar. It often does.

This is a reading of a published licence and not legal advice. Anyone building a business on these weights should put the actual text in front of a lawyer, because the definitions carry the weight and a paragraph on a marketing site is not where that decision belongs.

We check licence terms before model capability when we scope AI automation work, because a licence problem found after the build is a rewrite and a benchmark disappointment is a swap.