A form-filling AI the size of a photo just beat a hosted model

On Thursday @trycua introduced CUA-S1, a family of models it describes as "small, specialized, and built for computer use," and open-sourced the first one. CUA-S1-FORMS has 706,048 parameters. The whole checkpoint is 2.8 megabytes, smaller than most photos on your phone. A day earlier, @PrismML announced Ternary Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B that it says is "9x smaller than its full-precision counterpart" while keeping 98.2 percent of its benchmark score. The two releases have almost nothing in common except the point of them: capable AI that runs on hardware you already own.

A model that makes one kind of decision

CUA-S1-FORMS does not write anything. Give it a form field and a set of values already pulled from a document, and it returns a probability for each of four moves: use one of the values, check the box, click the button, or skip the element. It does this in a single pass. In the Hacker News launch thread the founders put the local decision at 7 to 9 milliseconds, against 260 to 280 milliseconds for a hosted call. The model sits behind Cua's browser driver as the decision layer, so ordinary code still does the typing and clicking.

Cua trained it in under 30 minutes on 10,000 synthetic forms. It scored 99.95 percent on a held-out synthetic test and got all 196 decisions right on three real forms. The comparison that got people talking: on Cua's own task it scored 99.7 percent against 83.6 percent for TypeSafe's hosted Jev, the "System One" decision model I wrote about last week. Cua borrowed the label and, on this narrow job, beat the original with a model you could send as an email attachment. The code, weights, and dataset are out under the MIT license.

A 27B model in 5.9 gigabytes

Bonsai 2 is the other end of the size range. A 27-billion-parameter model normally takes about 54 gigabytes at full precision. PrismML stores each weight as one of three values, minus one, zero, or plus one, which brings the file down to 5.9 gigabytes. Its write-up reports 143 tokens per second on an RTX 5090 and about 47 on an M5 Max Mac, with a 262K context window, image input, and an Apache 2.0 license. The footprint is unchanged from the first Bonsai in July, and the post says "the biggest change is quality," with the largest gains in agentic coding and tool use.

My read

The asterisks first. Cua's real-world test is three forms. Its model cannot invent a value, only choose among values something else already extracted, and the labels it learned are English. The model card says so. Bonsai's 98.2 percent comes from PrismML's own suite, and one independent local-model test log noted on Thursday that the first Bonsai's headline lead "is not supported at the suite level" in their agent runs, so they are waiting to see whether version 2 holds. AutoArk's report this week on streaming a 35B model from a Mac mini's SSD has the same shape: about 20 tokens per second in under 3 gigabytes of memory, paid for with a few benchmark points.

I still think the direction is right. For years the headline number in AI was how big a model could get. This week the more interesting number was how small one could get and still do the job. A 2.8 megabyte model has no per-call price and no rate limit, and nobody else's servers end up holding your customers' addresses. Moving data from one document into a form is the most repetitive step in most office workflows, and it turns out not to need a frontier model at all.

Cua's founders described the niche on Hacker News as the "space between a brittle script and a general agent loop," and that is the part I would remember. A script breaks the day the form changes, while a general agent is slow and expensive on the days it does not. A small trained model that only picks the next move sits between the two.

What this means for your business

Nobody needs to install a 27B model on the office laptop this week. The pattern is worth planning around, though:

• Separate the decisions from the writing. Whether an intake form is complete, or which invoice a payment belongs to, is a decision, and decisions are getting cheap and local fast.
• Keep the decision layer swappable. If your workflow calls a hosted model for every field today, put that behind one function so a smaller or local model can replace it without a rebuild.
• Ask where the data goes. A local model does not send your forms anywhere. A hosted one does, and its retention terms matter more than its benchmark.

If you want help finding which steps in your week are really repeated decisions, and what each one costs you now in time or API fees, that is what New Face Design's free process audit is for. Start here.