Are you using an open-weight Qwen image model for client work? Then something has changed since 20 September that you can't ignore. Alibaba's Qwen team released Qwen-Image-2.1, and in doing so swapped the familiar Apache 2.0 licence for the Qwen Research License Agreement: research and evaluation only, with commercial use requiring a separate agreement.
What's technically new, and what you give up for it
The model itself is a significant step. It is a 7B DiT with 32 single-stream layers, paired with a Qwen3-VL 8B text encoder and an 64-channel RGBA VAE. The last is the most striking: it generates and edits transparent images with a true alpha channel, and can cut a subject out of an ordinary photo as a separate layer. Beyond that: generation and editing in one checkpoint, up to ten reference images, local edits via masks or drawn circles, and native 2K output. For product photos, packshots and advertising material, that is exactly the kind of work a designer would otherwise bill you for.
The catch: the model card on Hugging Face now explicitly names the Qwen Research License Agreement, whereas the original Qwen-Image still falls under Apache 2.0. AI Weekly describes the switch and confirms that commercial users must arrange a separate licence.
This is more than a legal footnote. Our advice on Qwen has always been: the consumer app falls under Alibaba Cloud Singapore and Chinese law, so the sensible route is to self-host an open-weight model, that way your data stays within your own infrastructure. That mitigation still works for the language models, but no longer automatically for the newest image models. If you run Qwen-Image-2.1 locally for client campaigns, your GDPR story may be in order but your licence may not be. Check this if you work in marketing, e-commerce or communications and media; background is on our safety page.
In practice: stick with Qwen-Image (Apache 2.0) for commercial work, arrange a commercial licence, or switch. And record which model version is in your pipeline, "we use Qwen" is no longer an answer.
Step 5 Preview: 600 billion parameters for a dollar
On the same day, Chinese company StepFun launched Step 5 Preview: a sparse MoE with 600B parameters in total and 27B active per token, a context window of 1 million tokens, and image understanding. The price is the news: $1 per million input tokens and $2.70 per million output tokens, with a 95% cache discount. On Artificial Analysis's Intelligence Index it scores 44, level with Kimi K3 Max, according to AI Weekly. For comparison, that is a fraction of what Western frontier models cost per token.
Two caveats. First: the open weights were announced for 15 October, and for now only placeholders are on Hugging Face. So self-hosting isn't possible yet; your data goes to StepFun's own infrastructure in China for the time being. Second, "preview" means exactly that, don't rely on stable APIs or SLAs.
Interesting for experiments with long agent tasks, especially in software development. For personal data: don't use it until you know where it runs. The full comparison is on the models page.