{"schema":"sanctity-signal/0.2","record_type":"signal","id":"qwen-image-license-e028","domain":"models","drop_id":"signal-e028","occurred_at":"2026-09-20","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Alibaba's Qwen team","action_type":"licence terms changed","observed":"Sunday September 20, 2026. Alibaba's Qwen team published Qwen-Image-2.1 to Hugging Face, GitHub and ModelScope. The licence file on the Hugging Face model card, read this morning, is the Qwen Research License Agreement. The predecessor model, Qwen-Image, carries the line Qwen-Image is licensed under Apache 2.0 on its own model card, also read this morning. Under the new terms, commercial deployment requires a separate agreement with Alibaba; there is no monthly-active-user threshold of the kind other labs use, the restriction is on commercial use as such. The model: a 7-billion-parameter diffusion transformer of 32 single-stream layers, paired with an 8-billion-parameter Qwen3-VL text encoder and a 64-channel RGBA variational autoencoder, with two 9-billion-parameter prompt-rewriter checkpoints. Capabilities as published: native 2048 by 2048 generation at 40 steps, transparent-background output, editing with up to ten reference images, and mask, circle and painted-mark local edits. Alibaba's team states that architecture changes and key-value cache reuse speed up inference, particularly with multiple reference images, and claims the model beats most closed models on Qwen's own benchmark; no independent evaluation exists yet. It runs on a consumer graphics card such as an RTX 3090. The contrast inside the same weekend: StepFun published Step 5 Preview, a 600-billion-parameter mixture-of-experts model, through an API at 1 dollar per million input tokens and 2.70 dollars per million output tokens, and committed to releasing its weights openly on October 15.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"Hugging Face: Qwen/Qwen-Image-2.1 model card and Qwen Research License Agreement, read Sep 21","url":"https://huggingface.co/Qwen/Qwen-Image-2.1"},{"role":"primary","label":"Hugging Face: Qwen/Qwen-Image model card, licensed under Apache 2.0, read Sep 21","url":"https://huggingface.co/Qwen/Qwen-Image"},{"role":"corrob","label":"The Decoder: Alibaba's open-weight Qwen-Image-2.1 claims to beat closed models with 7 billion parameters, Sep 20","url":"https://the-decoder.com/alibabas-open-weight-qwen-image-2-1-claims-to-beat-closed-models-in-image-generation-with-just-7-billion-parameters/"},{"role":"corrob","label":"AI Weekly: Alibaba ships Qwen-Image-2.1, drops Apache for research-only licence, Sep 20","url":"https://aiweekly.co/alerts/alibaba-ships-qwen-image-21-7b-dit-with-native-2048x2048-and-rgba-drops-apache"}]},"strength":{"cost":"Low","commitment":"Medium","reversibility":"High","persistence":"Months","rationale":"Cost is low and we will not inflate it. Changing a licence file costs nothing to execute, and the revenue Alibaba forgoes by restricting commercial use is money it was never collecting. If anything the change is designed to create value rather than spend it. Commitment is medium: the licence is a published legal instrument attached to a shipped artefact, which is more than a statement, but it is one model and Alibaba has said nothing about the family. Reversibility is high, and this is the honest core of the card. A later release can go back to Apache 2.0 with no cost or explanation, and licences have been loosened before. Persistence is months rather than years for the same reason. The bars are modest on purpose: this is a small action with a large implication, and the size of the implication is not evidence about the size of the action."},"shift":{"prior":"Chinese labs competed on openness. Permissive weights were the instrument that put Qwen, DeepSeek and Kimi into Western products, and Apache 2.0 was the default.","action":"Sunday September 20, 2026. Alibaba's Qwen team published Qwen-Image-2.1 to Hugging Face, GitHub and ModelScope. The licence file on the Hugging Face model card, read this morning, is the Qwen Research License Agreement. The predecessor model, Qwen-Image, carries the line Qwen-Image is licensed under Apache 2.0 on its own model card, also read this morning. Under the new terms, commercial deployment requires a separate agreement with Alibaba; there is no monthly-active-user threshold of the kind other labs use, the restriction is on commercial use as such. The model: a 7-billion-parameter diffusion transformer of 32 single-stream layers, paired with an 8-billion-parameter Qwen3-VL text encoder and a 64-channel RGBA variational autoencoder, with two 9-billion-parameter prompt-rewriter checkpoints. Capabilities as published: native 2048 by 2048 generation at 40 steps, transparent-background output, editing with up to ten reference images, and mask, circle and painted-mark local edits. Alibaba's team states that architecture changes and key-value cache reuse speed up inference, particularly with multiple reference images, and claims the model beats most closed models on Qwen's own benchmark; no independent evaluation exists yet. It runs on a consumer graphics card such as an RTX 3090. The contrast inside the same weekend: StepFun published Step 5 Preview, a 600-billion-parameter mixture-of-experts model, through an API at 1 dollar per million input tokens and 2.70 dollars per million output tokens, and committed to releasing its weights openly on October 15.","update":"Read each Chinese release's licence file as a separate decision rather than a house policy, and watch whether the research licence spreads from images to the language models that carry the strategy."},"inference":{"established":["Qwen-Image-2.1 was published on September 20, 2026 to Hugging Face, GitHub and ModelScope by Alibaba's Qwen team.","Its licence, as shown on the Hugging Face model card, is the Qwen Research License Agreement, and commercial use requires a separate agreement.","The predecessor model Qwen-Image is licensed under Apache 2.0, stated on its own model card.","The model is a 7-billion-parameter diffusion transformer with an 8-billion-parameter text encoder, native 2048 by 2048 output, transparency support and editing with up to ten reference images."],"unknown":["That Alibaba is abandoning open weights. Its language models are unchanged and no policy statement accompanied the release.","That Chinese open releases are ending. StepFun committed the same weekend to publishing 600-billion-parameter weights on October 15.","That the model is as good as claimed. The benchmark cited is Qwen's own and no independent evaluation has been published.","That anyone will be stopped. A research licence changes what is lawful, not what is technically possible, and enforcement across jurisdictions is a separate question entirely."]},"narrative":{"hook":"Alibaba released a new image model and took away the right to use it commercially.","plain":"On September 20 Alibaba's Qwen team published Qwen-Image-2.1, a 7-billion-parameter image generator with native 2048 by 2048 output and transparent-background support, and released the weights under the Qwen Research License Agreement instead of the Apache 2.0 licence that covered the previous Qwen-Image. Anyone may download it; only researchers may use it, and companies now need a separate commercial agreement from Alibaba. The caveat that matters is scope: this is one model in one family, and Alibaba's language models remain under their existing terms, so this is a change of policy on a single release rather than an announced retreat from open weights.","story":["For two years the strategy has been legible. Chinese labs could not win on closed frontier models against companies with more compute, so they gave the weights away under licences with no conditions attached, and the result was that Qwen, DeepSeek and Kimi ended up inside Western products, on Western developers' machines, and in the training pipelines of American startups. Edition 018 carded an American company shipping its main coding model on top of a Chinese one. Edition 021 carded a Chinese state laboratory releasing a 397-billion-parameter model with no conditions at all. Apache 2.0 was not generosity, it was distribution. On September 20 the lab that did most to establish that pattern shipped its best image model and attached a licence that says researchers only.","The restraint worth keeping is that this is one model. Qwen-Image-2.1 is an image generator, the part of the field where the commercial value is most immediate and most easily captured by a competitor who did not pay to train it, and Alibaba's language models have not moved. It is also the weekend StepFun went the other way, putting a 600-billion-parameter model behind a cheap API and promising the weights on October 15, which is the pattern holding. But a licence file is an action in a way that a strategy essay is not: it is enforceable, it is dated, and it changes what a company in Hamburg or Bangalore may legally ship next week. Read each Chinese release's licence file as a separate decision rather than a house policy, and watch whether the research licence spreads from images to the language models that carry the strategy."],"trail":[{"d":"AUG 2025","t":"Alibaba publishes Qwen-Image under Apache 2.0, with no restriction on commercial use."},{"d":"SEP 10","t":"Cognition ships its main coding model built on a Chinese open-weight model. Carded in Edition 018."},{"d":"SEP 13","t":"Shanghai Artificial Intelligence Laboratory releases a 397-billion-parameter model with no conditions attached. Carded in Edition 021."},{"d":"SEP 20","t":"Alibaba publishes Qwen-Image-2.1 under the Qwen Research License Agreement, with commercial use requiring a separate deal."},{"d":"Reacts","r":true,"t":"Every startup and agency that built an image pipeline on Qwen-Image and now has to check which version it is running; Black Forest Labs, Stability and the other open image model houses, whose main free competitor just became conditional; Hugging Face and ModelScope, which host the weights and the licence text together; Adobe and Canva, whose paid products competed against a free one; American policymakers who have argued Chinese open weights are a strategic giveaway; Chinese rivals including StepFun, which promised open weights on October 15 two days earlier; and the EU AI Act's open-source exemptions, which turn on exactly this kind of licence wording."}],"glossary":[{"term":"Apache 2.0","def":"A permissive licence letting anyone use, modify and sell software built on the work, with almost no conditions. It is the licence most open-weight model releases have used."},{"term":"Research licence","def":"Weights anyone may download and study, but only non-commercially. Building a product on them requires a separate paid or negotiated agreement with the owner."},{"term":"Open weights","def":"Publishing the trained numbers inside a model so others can run it themselves. It is not the same as open source, and the licence attached decides what may actually be done with it."},{"term":"RGBA","def":"An image format that carries transparency alongside colour, so a generated object can be placed over another background without a cut-out step."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"The open-weight strategy has started to close.","text":"The lab that made permissive Chinese weights a fact of the industry attached a commercial restriction to its best model. Distribution has done its work and the licence is where the turn shows first.","abstain":false},{"id":"b","label":"A product decision about images, not a policy about models.","text":"Image generation is the one area where a free release is immediately monetisable by someone else. Restricting it while leaving the language models open is a pricing choice, not a strategic retreat.","abstain":false},{"id":"c","label":"A move against specific competitors.","text":"A research licence blocks rivals from shipping the model in a paid product while leaving the goodwill of a public release intact. It is aimed at the companies that would have taken it for free.","abstain":false},{"id":"d","label":"The normal maturing of any open project.","text":"Almost every widely adopted open release eventually adds conditions once it has users worth conditioning. Qwen doing it in its second year is the ordinary arc, not a signal about China.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"Alibaba publishes a Qwen language model of 30 billion parameters or more under a research or otherwise non-commercial licence, or states a change of licensing policy for the Qwen family, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"Alibaba publishes its next major Qwen image or language model under Apache 2.0 or another licence permitting commercial use, or relicenses Qwen-Image-2.1 permissively, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"}]}