A 600B model for $1, and 'open weights' you can't sell with

Early Saturday, @StepFun_ai introduced Step 5 Preview, a flagship "for agentic work, delivering frontier-level performance across software engineering and professional knowledge work." Ten hours later, @Alibaba_Qwen announced Qwen-Image-2.1 and led with three words: "Now open weights!"

Both announcements use the word open. They mean very different things by it, and the difference decides whether you can legally build anything you sell on either one.

The cheap one

Step 5 Preview is a 600 billion parameter sparse mixture-of-experts model with 27 billion parameters active per token and a one million token context window. API access opened the same day. Artificial Analysis scores it 44 on its Intelligence Index, 24th of 200 models, at $1.00 per million input tokens and $2.70 per million output. The median score for reasoning models in that price bracket is 24, and Step 5 runs close to 100 output tokens per second.

The price is what makes this worth reading about. Forty-four is nowhere near the top of the board. But agentic work burns tokens the way a delivery van burns gas, with a long context window loaded on every call and a fresh tool call each time a step fails and has to run again. At $2.70 per million output tokens, jobs that were too expensive to hand to a model last year stop being expensive.

StepFun says the weights open on October 15. They are not out yet, and the terms that will govern them have not been published. Artificial Analysis still lists the model as proprietary with weights unavailable. So the open part of this one is a promise with a date on it.

The open one with a gate

Qwen's release is the more interesting case, because the weights really are downloadable right now. Seven billion parameters in the visual generation component, text-to-image and editing in one model, transparent layer output, up to ten reference images for composition.

Then you open the license file. It is the Qwen Research License Agreement, dated September 20, 2026, and it defines "Non-Commercial" as "for research or evaluation purposes only." The Hugging Face model card lists the license as qwen-research. The original Qwen-Image shipped under Apache 2.0, which let you download it, fine-tune it, and put it in a product you sell. This one routes commercial use to a separate agreement Alibaba has not priced publicly.

So a designer at a St. Charles shop can run Qwen-Image-2.1 on their own machine tonight and should. They cannot put its output on a client's packaging without a conversation with Alibaba's legal team.

My read

"Open weights" now covers at least three different deals:

• Weights released under a permissive license like MIT or Apache 2.0, where you can build a business on it
• Weights released under a research license, where you can inspect and run it but not sell with it
• Weights promised on a future date, with terms to be announced

Headlines and X posts flatten all three into the same two words, which is how people end up building on terms nobody read. The two releases above landed hours apart in different categories, and going by the announcements alone you would file them together.

I do not read this as bad faith. Training a frontier model costs a great deal, and labs are working out how to keep funding it while staying visible to developers. A research license is a reasonable middle position. The practical effect is still that the license file now decides as much as the benchmark does, and it can change between versions of the same model line, from the same team, without the name changing much.

What this means for your business

If you or a contractor is about to build something on a downloadable model, the license file is the first thing to read, not the last. It takes two minutes and answers the question you actually care about: can I sell work made with this. Check it again on every upgrade, because Qwen just showed that the answer moves.

The pricing side cuts the other way, and in your favor. Vendors have been telling owners for two years that AI features cost what they cost. A model scoring above the median for its price bracket while charging $2.70 per million output tokens is a reminder that the floor is still dropping, and that any quote you accepted six months ago deserves a second look. The cheapest model that clears the bar for your specific task is rarely the one in the headlines.

If you want a straight answer on which parts of your week a model could handle, what each one would actually cost to run, and what the license lets you do with the result, that is what New Face Design's free process audit is for. Start here.