Cua open-sources a 2.8 MB model for computer-use form filling

An MIT model that scores each form action in one pass, weights and dataset on Hugging Face — plus a 2-bit 27B you can run locally in ~6 GB.

Nowline SEP 20 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • 706,048 parameters, and it fills the form

    CUA-S1-FORMS is out under MIT: it scores each form decision — fill a field, check a box, click, or skip — in a single forward pass, then hands the ordered actions to Cua Driver to execute. The weights are a 2.8 MB checkpoint; training code and the 234,067-row dataset ship alongside it on Hugging Face.

  • 99.7% vs Jev's 83.6% — with an asterisk

    On Cua's own form task it beats TypeSafe's managed Jev (open-sourced four days earlier) 99.7% to 83.6%, and hits 99.95% top-1 on synthetic data. But the real-form number — a perfect 196/196 — comes from just three forms and three PDFs, so treat it as a proof of concept, not a guarantee.

  • What to build this weekend

    Drop it into a computer-use agent to auto-fill web forms and PDFs at effectively zero inference cost — it runs locally and generates no tokens. The catch: it only selects from values you already extracted, so it won't invent a missing field. Pair it with a solid document extractor and let the tiny scorer drive the clicks.

  • Elsewhere: a 27B model that fits in ~6 GB

    Prism ML's Ternary Bonsai 2 27B uses 2-bit ternary quantization to squeeze a 27B model into roughly 5.9–8.6 GB while claiming ~98% of the original's quality. It's live on OpenRouter and Hugging Face (GGUF and MLX builds), so a mid-range laptop can now host a model this size.