Qwen3.8-27B License: Planning Commercial Use Before the License File Exists

You cannot read a license that does not exist yet, and for Qwen3.8-27B that is exactly where things stand. Alibaba announced the ~27B open-weights model on August 3, 2026 alongside Qwen3.8-Max, promised the weight drop "within about a week," and as of this writing has published no repository and no LICENSE file. For a commercial team, that gap is the story: the license decision you have to make — what you can build, what you can ship, and where you will need Alibaba's sign-off — has to be made on the basis of the two licenses that have historically governed Qwe​n weights, and then checked against the file the day it lands.

This piece walks through what is confirmed, the two licenses Qwe​n has used, what the 100-million-MAU clause actually says, and a pre-drop checklist so the license is not the thing that blocks your launch.

What is confirmed, and what is not

Confirmed: Qwen3.8-27B is announced as an open-weights release, planned for Hugging Face and ModelScope, and it is the self-hostable member of the Qwe​n3.8 generation. That is nearly all.

Not confirmed: the license itself. The generation's other flagship, Qwen3.8-Max, is an API model and does not ship weights at all, so it sets no precedent for the 27B. Some outlets have already written that the 27B ships under Apache 2.0 — treat those as speculation, not reporting. Alibaba has made no license statement, and the pattern across Qwe​n history is exactly why you cannot assume.

The two licenses that have governed Qwe​n weights

Qwe​n separates source code from model weights, and the two have different licenses. The code in the QwenLM/Qwe​n GitHub repository — demo, fine-tuning, and serving code — has shipped under Apache 2.0. The model weights are licensed per release, and Qwe​n has used two different agreements:

• Apache 2.0 — permissive. Unrestricted commercial use, no user cap, no royalty, you may fine-tune and redistribute with the license and notices attached, and it carries an explicit patent grant. This is what the Qwe​n3 line (including the 235B flagship) and most small and mid-size Qwe​n models shipped under.

• Tongyi Qianwen License Agreement — conditional. Commercial use is permitted, but only for organizations under 100 million monthly active users. Above that threshold you must request a license from Alibaba. It also restricts building competitive AI services, requires attribution and license-notice preservation, and is governed by Chinese law with Hangzhou jurisdiction.

For Qwen3.8-27B, either outcome is on the table until the repository exists. Apache 2.0 is the likelier default given the last two generations, but "likely" is not a compliance posture.

The 100-million-MAU clause, decoded

The Tongyi Qianwen threshold is the clause that matters most, and it is worth reading literally. The agreement says that if you are commercially using the materials and your product or service has more than 100 million monthly active users, you shall request a license from Alibaba. Note three things:

• The threshold is 100 million MAU — seven times lower than Meta's 700-million-user threshold on Llama 2 and Llama 3. It applies to a meaningful slice of production deployments, not just hyperscale apps.

• It is a "request a license" trigger, not an automatic grant. The negotiation outcome is Alibaba's to decide, on terms you do not control.

• The competitive-AI restriction is separate from the MAU cap and applies regardless of size: building a service that competes with Alibaba's own AI offerings can be blocked even well below 100M MAU.

For most teams — the SaaS product at tens of thousands of users — the MAU cap never binds. But a large installed base, or a product that looks like a general-purpose AI assistant, changes the calculus.

What each license means for your product

Under Apache 2.0, the Qwen3.8-27B becomes a boring, safe dependency: fine-tune it, embed it, white-label it, serve it, and redistribute derived weights so long as you keep the license and notices and do not use Alibaba's trademarks. No MAU counting, no negotiation, no jurisdiction surprises. This is the outcome most open-model product teams are hoping for, and it is the reason the license question has real commercial weight rather than being a formality.

Under the Tongyi Qianwen license, the same actions are fine up to 100M MAU, then the permission requirement kicks in, and the competitive-AI clause narrows what you can build. Attribution and acceptable-use conditions apply at every scale. The license's Chinese-law and Hangzhou-jurisdiction clauses are worth flagging to counsel even if you never cross the threshold — it is a materially different legal posture from Apache 2.0 for the same deployment tier.

The pattern across generations — and why 3.8 is a new decision

Qwe​n's licensing has trended toward Apache: the Qwe​n3 generation shipped even its 235B flagship under Apache 2.0, and Qwe​n3.5 and Qwe​n3.6 releases followed the permissive path. The Tongyi Qianwen license has been reserved for older and larger items — Qwen2-72B, Qwen2.5-72B, the VL-72B variants, and embedding/reranker models. If past is precedent, the 27B lands Apache. But the 3.8 generation is also the one where Alibaba is pushing agentic autonomy and enterprise positioning harder than before, and a generation's flagships have changed license direction before in this industry. The precedent is a bet, not a contract.

The hosted-API escape hatch

There is a wrinkle worth separating early: the model-use license governs the weights — who may distribute, modify, and run them. It does not govern every way you touch the model. If you call the model through a hosted API rather than downloading the weights, the provider's API terms govern your use, and the weights-license triggers (redistribution, MAU thresholds, fine-tuning rights) are the provider's burden, not yours. That is the clean path for teams that want Qwe​n-class capability without becoming a licensee.

That distinction is why the API route matters for this specific model while it is still unproven. Qwen3.8-Max — the flagship of the same generation — is already callable through OrcaRouter at the provider list price, $2 per million input tokens and $6 per million output, passed through at zero markup, with automatic failover across providers. You get the generation's capability on the vendor's API terms, with no weights to license and no download to vet, while the 27B's license and quality sort themselves out in public. When the 27B eventually ships and a provider hosts it, your use of that hosted endpoint is that provider's API terms as well — worth remembering before you assume a permissive weights license is required for every deployment.

A pre-drop checklist for commercial teams

• Confirm the repository is on the official Qwe​n Hugging Face or ModelScope organization before trusting any download

• Read the actual LICENSE file — not the blog post, not a press summary, the file in the repo

• Determine which of the two licenses it is, and have counsel confirm the MAU clause and competitive-AI restriction apply or not

• Count your MAU against the threshold if you are anywhere near the 100M line — do not guess

• If the license is Tongyi and you plan to fine-tune and redistribute derived weights, map the attribution and notice requirements into your distribution process now

• If you are not sure you want to be a weights licensee at all, route the model through a hosted API instead and let the provider carry the license

The bottom line: Qwen3.8-27B's license is a live decision with two plausible answers, and neither is decided yet. Apache 2.0 would make it an uncomplicated commercial dependency. The Tongyi Qianwen license would make it fine for the vast majority of teams — under 100M MAU — with specific obligations above that line. Plan for both, read the file the day it lands, and keep the hosted-API route open if you would rather not be a licensee at all. The license file is not a detail you discover after you ship; it is a decision you make in advance.

FAQ

If I call Qwen3.8-27B through a hosted API, does the 100M MAU cap apply to me?

No — the weights license binds whoever distributes or modifies the weights. A hosted API provider's terms govern your API use, so the MAU threshold and redistribution rules sit with the provider, not your application.

Is Apache 2.0 a safe assumption for Qwen3.8-27B?

No. The last two generations shipped Apache 2.0, but Alibaba has made no license statement for the 27B, and Qwe​n has used the more restrictive Tongyi Qianwen license for prior models. Assume nothing until the LICENSE file exists.

What happens if my product crosses 100 million MAU under the Tongyi Qianwen license?

You must request a license from Alibaba before continuing commercial use. There is no automatic grant, and the competitive-AI restriction applies regardless of your size.