AI Brief, 21 September 2026: an image model that keeps its alpha channel, and gives up its licence

Alibaba published Qwen-Image-2.1 on Saturday 20 September, and it is the only substantial set of open weights released in the window. The model card went public at 09:41 UTC, the licence file carries a release date of 20 September, and the demo went up thirty-four minutes later. It is a 16.2-billion-parameter image generation and editing stack that does something no other current open release does: it treats transparency as a first-class output, carrying an alpha channel end to end through the autoencoder rather than bolting on a matting step afterwards.

Two things about it deserve more attention than the capability. The first is the licence. Qwen-Image 1.0 was Apache-2.0. Qwen-Image-2.1 is released under the Qwen Research License, which grants rights "for non-commercial purposes only" and defines non-commercial as research or evaluation. Anyone building a product on the previous version and expecting to upgrade is looking at a licensing conversation with Hangzhou Tongyi Laboratory, not a version bump. This is a meaningful narrowing of what "open weights" bought a week ago, and it was not announced as a change.

The second is that the release contains no benchmark numbers at all. Not a table, not a score, not a comparison, in either the model card or the project's GitHub README. There is no technical report and no preprint. Figures attributed to the lab are circulating in secondary coverage, and they are scores on Qwen-Image-Bench, a benchmark Qwen itself authored and published in May. None of them could be traced to a primary source, so none are reproduced here. What does exist, eight hours after launch, is one practitioner's manual scoring and two independent measurements of speed and memory.

Separately, and a day before this window opened, Artificial Analysis changed how its Intelligence Index is computed for the fourth time in nineteen days. It moved the anchor of one component's Elo scale onto a specific mid-table open-weights model, and 145 models' index scores moved. The change shipped with no announcement article. This account has asked for sixteen consecutive issues whether scores from different index versions can be compared; the answer is still nothing.

• Qwen-Image-2.1 released 20 September: 16.22B parameters total, 33.1 GB download, weights ungated, licence changed from Apache-2.0 to a non-commercial research licence.
• Its autoencoder is genuinely four-channel — alpha travels the same trunk as RGB into one 64-dimensional latent. The text encoder never sees it.
• No benchmark figures appear anywhere in the release. The only independent scoring so far is one evaluator's manual run: 7 of 15 prompts passed, against 4 of 15 for the previous version.
• Artificial Analysis Intelligence Index v4.3.2, dated 19 September, pinned its GDPval-AA Elo scale to DeepSeek V4.1 Flash at 1600 and refitted two evaluations. 145 index scores changed; 128 fell.
• Treasury Secretary Scott Bessent said on 20 September that the US proposed an incident-notification mechanism for AI to China. No government on either side has published anything about it.

Qwen keeps the alpha channel, and takes back the licence #

The repository was created on 14 September and the first weights landed on the 15th, so the public release on the 20th ends a week of private staging rather than a same-day upload. One detail in that history is worth stating precisely, because the commit message overstates it. On 19 September a commit titled "Update to 0919 release checkpoint" replaced the two transformer shards and nothing else: the four text-encoder shards and the autoencoder are bit-for-bit the files uploaded on the 15th. The same commit claims to fix the autoencoder's spatial scale factor to 16, but that was an edit to a configuration file, not to a weight — the value had been recorded as 8 and the tensors were always those of a 16× autoencoder.

The parameter split is where the interesting structural claim sits, and it is not what the repository's own metadata reports.

The text encoder is a Qwen3-VL vision-language model at 8.77 billion parameters, larger than the 7.12-billion-parameter diffusion transformer it conditions. The autoencoder is small, 338 million parameters, but stored in float32, so it still occupies 1.35 GB. The card's own framing — "7B parameters in its visual generation component" — is exact; it is the repository's generated parameter badge that misleads, reporting the transformer alone.

How the transparency actually works. There is no alpha head and no separate decoder. The autoencoder's first convolution has shape [96, 4, 3, 3] and its last has shape [4, 144, 3, 3]: four channels in, four channels out, with alpha compressed into the same 64-dimensional latent as colour. The asymmetry is on the conditioning side. When a reference image carries transparency, the autoencoder reads all four channels, but the copy handed to the vision-language encoder is first composited over white — the checkpoint was trained that way. So the generator reasons about alpha while the thing interpreting your prompt does not.

Multi-reference conditioning, up to ten images, runs along two paths at once. Each reference gets a literal index token in the prompt template, <image1>, <image2>, which is why the companion prompt-rewriter models emit instructions like placing one image's subject into another's scene. Spatially, every reference is encoded, flattened and concatenated onto the sequence ahead of the noisy target, and the separation between them is enforced by an attention mask rather than by architecture: causal across blocks, bidirectional within each image.

Because the reference and text tokens are modulated from the first step, their keys and values do not change as denoising proceeds, so they are computed once and cached. That is the mechanism behind the claim that the speedup is largest when you pass many references.

Transparency is not a flag. It is steered by prompt wording, and the card's guidance is to say so explicitly. Note also that the pipeline defaults to 1024 pixels unless you pass dimensions, and that it is designed to be sampled without classifier-free guidance:

from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1", torch_dtype="bfloat16").to("cuda")
img = pipe(
    prompt="A cartoon dragon sticker. The image has an alpha channel and the background is transparent.",
    width=2048, height=2048,     # without these you get 1024, not the card's example size
    true_cfg_scale=1.0,          # the checkpoint is meant to be sampled without guidance
).images[0]
img.save("dragon.png")           # four channels; alpha survives the round trip

What is actually measured. One evaluator running a manual fifteen-prompt comparison scored it 7 of 15, against 4 of 15 for Qwen-Image 1.0, and called it "a pretty big leap over the last open-weight version" while noting visible signs of synthetic training data in some outputs. That is one person's judgement on a small prompt set, and their comparison figures for rival models shifted between posts, so treat it as a signal rather than a score. Two harder numbers came out of the same discussion: roughly five seconds for a one-megapixel image on an RTX 4090 at int8, and a runtime footprint of 15,645 MB at Q8 — 7,670 MB of it the text encoder, 7,331 MB the diffusion model. A 16 GB card is therefore marginal, an unusual position for a model whose generator is only 7B.

Artificial Analysis moved its anchor, and 145 scores moved with it #

On 19 September, a day before this window opened and unremarked here yesterday, Artificial Analysis published Intelligence Index v4.3.2. It is a changelog line with no accompanying article, and it does two things at once: it pins the Elo scale of its GDPval-AA evaluation to DeepSeek V4.1 Flash at 1600, and it refits both GDPval-AA and AA-Briefcase using Crowd-BT, a Bradley-Terry variant with a per-annotator reliability term, which the firm cites properly.

Why an Elo scale needs an anchor at all is the part worth understanding. A Bradley-Terry fit recovers only differences in strength from win-loss records: adding a constant to every rating leaves the likelihood unchanged, so the fit is under-determined by exactly one degree of freedom and something must be pinned. Until now the pin was human expert performance, set at 1000 in June. It is now a specific model, and the human reference point has disappeared from the current methodology section.

The conversion into the index is published, and it makes the stakes arithmetic. A model's contribution is $\mathrm{clamp}((E-500)/2000)$ , where $E$ is its fitted Elo, and that contribution carries 10% weight. So 1 index point equals 200 Elo. Pinning DeepSeek V4.1 Flash at $E=1600$ gives $(1600-500)/2000=0.55$ exactly, and that is precisely the value its published component now shows — the anchor is live and checkable. Its previous value was 0.566, from an Elo of 1632.

Artificial Analysis published no before-and-after table and named no affected model. The figures that follow come from comparing archived daily captures of its own leaderboard from either side of the change, not from the firm.

Across the 233 models carrying a GDPval-AA rating on both days, the mean Elo fell 60.7 and the median fell 59.1: 177 fell, four rose, and the 52 that held steady were already clamped at the floor. The anchor itself moved down only 32, so this is not a rigid translation of the scale — the fit compressed the lower half, and the number of models pinned to the floor rose from 52 to 74. Fifteen models were demoted from measured to estimated; none moved the other way.

At the index level 145 scores changed, 128 of them downward. Restricting to the 129 models the firm flags as measured rather than estimated, the mean change was −0.36 and 126 of them fell. Fifty-six models' displayed, rounded score changed, including GPT-6 Astra at extra-high effort, which went from a displayed 53 to 52.

The top of the table is where this closes a loop. On 7 September this brief recorded Claude Fable 5.1 leading GPT-6 Astra by 0.56 index points under v4.3, with both displaying 53. That figure was still exactly 0.5597 on 19 September. It is now 0.6813. The gap widened by 0.12 points, and no new evaluation results were involved: one lab gained relative ground on a change of fitting method.

Attributing that to the anchor alone would be wrong, and it is worth saying why, because it is the cleanest illustration of what shipped. AA-Briefcase carries 15% weight, more than GDPval-AA's 10%, and its per-model values are published nowhere. Claude Opus 5 at maximum effort saw its GDPval Elo fall 27.2 — worth about −0.14 index points — yet its index score rose by 0.08. Something in the Briefcase refit moved it roughly two tenths the other way, and no public data can isolate it.

Four index versions have now shipped in nineteen days: v4.2 on 4 September, v4.3 on the 7th, v4.3.1 on an unstated date, and v4.3.2 on the 19th. Only two of the four have a changelog entry at all; v4.3.1 exists solely as a line in the methodology page's version history, recording that the judge panels were refreshed. The rescored data reached the public leaderboard on 19 September itself, between 03:02 and 22:57 UTC.

And the standing question, asked here since 5 September, still has no answer. Searching the changelog, both methodology pages and the v4.2 and v4.3 announcement articles for any statement about whether index versions can be compared returns nothing. The firm is plainly capable of saying it: the same methodology page says "not directly comparable" four separate times about individual evaluations, including its own AA-LCR versions. It has never once said it about the index.

There is a sharper version of the problem now, and it is checkable. The v4.3 announcement article, dated 7 September and still carrying that byline, is being re-rendered live under the current version. Its opening sentence now reads "Intelligence Index v4.3.2 incorporates 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1"; archived copies from 7, 13 and 17 September all read v4.3, AA-Briefcase, GDPval-AA v2. The prose beneath is frozen at publication and still states that both leading models score 53 on v4.3. A reader arriving at a dated article gets a twelve-day-old sentence next to current numbers, with no edit note.

Washington proposes telling Beijing about AI incidents #

Speaking to reporters on Sunday 20 September, after roughly eight hours of trade talks with Chinese Vice Premier He Lifeng in New York, Treasury Secretary Scott Bessent said the United States had proposed a mechanism for notifying China about AI incidents that rise to a national-security level, within a body he called the US-China AI dialogue. Trade Representative Jamieson Greer was present.

This is a genuinely interesting governance idea — the state-to-state analogue of a nuclear hotline, applied to model behaviour — and almost nothing about it exists in writing. There is no Treasury readout, no filing from the Trade Representative, no joint statement, and the Chinese state-media account of the talks acknowledges only "issues relating to AI" without specifics. Bessent's characterisation that both sides agreed to the dialogue is therefore one side's account of a bilateral understanding, reported speech rather than a commitment either government has published. What it would cover, what threshold "national security level" means, and whether it binds anyone are all unstated.

Also notable #

• Trump said on Truth Social on Saturday 19 September that he is forming an "AI Force", likening it to his own creation of Space Force, and will name an AI czar "in the near future". The artifact is a social-media post and nothing else: no executive order appears in the White House presidential-actions listing, and there is no nominee, agency, budget line or statutory vehicle. Space Force was created by Congress in the FY2020 defence authorisation, so the comparison does not describe a mechanism available to a president alone. The role is also being re-created rather than created — David Sacks held it until his 130-day special-government-employee term expired in March.
• Google's ax agent orchestrator reached the Hacker News front page late on 20 September with over 330 points, and much of the discussion treated it as new. It is not: it sits in the verified google GitHub organisation under Apache-2.0, but its first release was 20 May and Sunday's v0.3.0 is a maintenance commit. Worth noting that it builds on a substrate library in a non-Google organisation that does carry the "not an officially supported Google product" disclaimer.
• Runway showed real-time steerable video generation in a post on 20 September, described as frame-by-frame autoregressive generation with drift handled by training on the model's own outputs. There is no paper, no weights, no product and no timeline, and the claim rests on a single company post relayed by one outlet. The sub-100-millisecond first-frame figure circulating alongside it belongs to a different research preview from March.
• The two prompt-rewriter models shipped beside Qwen-Image-2.1 were published at 09:45 UTC with quickstart instructions pointing at a system-prompt file that was not added to either repository until 11:59.

What to watch #

• Whether an independent evaluation of Qwen-Image-2.1 appears with a published harness. The only scoring so far is one evaluator's manual pass on fifteen prompts, and the lab's own figures are on a benchmark the lab wrote. The editing half has not been run by anyone.
• Whether Qwen states why the licence changed, and whether the research licence is now the default for the Qwen-Image line or specific to this checkpoint. The move from Apache-2.0 went unremarked in the release, and it decides whether the previous version stays the one people build on.
• Whether Artificial Analysis says anything about cross-version comparability, asked here for a seventeenth consecutive issue and now with a concrete case attached: a dated article whose version label updates while its sentences do not. Republishing old prose under new numbers is the strongest argument yet that the position needs stating.
• Whether either government publishes anything on the AI incident-notification proposal. A single readout naming a threshold would turn this from reported speech into policy.
• Independent numbers for Tencent's Hy4 preview are absent from Artificial Analysis for a sixteenth consecutive issue. The model does not appear in its data at all.