A model you download makes transparent images
Edition of 2026-09-21
# AI For All ### Monday, September 21, 2026
As of today, a designer can download a model that generates and edits transparent images, and keep it running on a machine they own, instead of paying per picture.
The picture tool you keep instead of rent
On Sunday the Qwen team published Qwen-Image-2.1. One model does both jobs: it makes an image from a written description, and it edits one you already have. The part that is genuinely new is transparency. It produces cut-outs directly, edits the separate layers, and lifts a subject out of a photograph without a second tool doing the matting. In a single pass it will work from up to *10 reference images*, which is what matters when a layout is assembled from several sources rather than described from scratch.
Availability is the other half. The weights landed on Hugging Face and ModelScope the same day, and Diffusers, ComfyUI, vLLM-Omni and SGLang all added support on day one. That is the difference between a repository and a tool: if the software a designer already opens can load it, the model is usable on Monday morning.
The reserve is the licence, and it is not a small one. It is a research licence, so a business cannot put the output to commercial use. Anyone paying a per-image service today and reading this as a cheaper replacement should read the licence first.
Read the release at Qwen on GitHub.
The market that had three suppliers now has a fourth
On Sunday, at the World Manufacturing Convention in Hefei, ChangXin Memory Technologies said its fifth-generation memory platform is in mass production, and showed two low-power memory products built on it for mid-range and high-end phones. The company's own numbers are the argument: a single chip holds *3 gigabytes, half again as much as the previous generation, and one wafer now yields at least 50%* more chips than the platform it replaces. That second number is the one to hold on to. Dies per wafer is the lever that sets the cost of a chip, and cost is the only thing that has ever moved this market.
The context is what makes it a story rather than a specification. Leading-edge memory has had three suppliers, Samsung Electronics, SK hynix and Micron, and their pricing power rests on that short list. CXMT's vice president, Luo Xiaodong, described the company's process capability as on par with the most advanced nodes in volume production.
Two reserves, both from the company's own account of the facts. It did not disclose yields or monthly output, and dies per wafer counts total dies before defects are screened out, which is not the same as the number that ships. Analysts also note that the process measurement the company published describes one structure inside the chip, and does not map cleanly onto the node names its three rivals use. Anyone holding memory exposure should read the caveat as carefully as the headline.
Read the report at Seoul Economic Daily.
The ad code that reports back who visited
A teardown published on Sunday describes a mechanism that turns an ordinary advertising pixel into a link between two identities. The author says he reproduced it on his own phone, with two independent methods of capturing the traffic. When someone uses ChatGPT, a cookie is set on OpenAI's advertising collector and bound to that account. When that same person later visits a site that has bought advertising inside ChatGPT, the site runs a piece of OpenAI's code, and the code sends the cookie back along with data about the page being viewed. Put the two together and OpenAI can join a browsing session to an account. He reports the scale he counted while cross-checking months of observed traffic: *936 distinct advertiser pixels across 1,029 hostnames*.
The concrete part is the direction of the data. A company that buys ads on ChatGPT installs the code on its own site, exactly as retailers already install tracking code from other platforms, and that is what does the reporting. Nobody has to visit OpenAI to be measured.
The reserve is heavy, and it is why this is a finding and not a product review. It is one researcher's account. OpenAI has not responded. The cookie policy it would be checked against sits on a domain this server cannot reach, and no independent measurement confirms the mechanism. Read it as the most concrete description yet of how the advertising build connects a browsing history to an account, and not as something OpenAI has confirmed.
Read the teardown at Buchodi.
Weak signal
A federal antitrust suit filed on Friday claims that Anthropic, OpenAI, SpaceXAI and Google agreed among themselves to slow their own progress, and that the agreement reduces what subscribers get for the same money. The plaintiffs are paying subscribers who buy ChatGPT, Claude, Grok or Gemini, suing on behalf of a proposed nationwide class of other paid users. The complaint points to the day in September when Anthropic's chief executive, Dario Amodei, published an essay urging the industry to pace the frontier, and to the same-day endorsements from Sam Altman, Elon Musk and Demis Hassabis. The companies did not respond to requests for comment, and no court has examined the agreement described.
This is the biggest story of the weekend, and it is here rather than above for a mechanical reason that is worth stating plainly. The wire copy carries no figure in the form this publication is allowed to verify, which is a number joined to a unit. Its quantities are a count of plaintiffs written as a word, a date, and a docket number. Rather than print a number that has the look of something checked, we print none, and you get the claim without it.
The report is carried by the Associated Press.
For anyone whose next purchase is a picture tool, the licence is the whole story.