Every week there's a new frontier model with a trillion parameters and a price tag to match. So when a startup open-sources an AI model small enough to fit in a WhatsApp attachment, my ears perk up. On September 18, Francesco Bonacci — founder of Cua, the San Francisco company behind one of the most-starred open-source computer-use agent projects on GitHub — released CUA-S1-FORMS. Its entire job? Deciding how an AI agent should fill out forms. That's it. And that's exactly why it's clever.
01.What Cua actually released
CUA-S1-FORMS is the first release in what Cua calls its CUA-S1 family of specialised "System One" models. The numbers are almost funny by 2026 standards: 706,048 trainable parameters and a 2.8 MB checkpoint. For perspective, that's roughly a thousand times smaller than the compact models you'd normally run on a laptop — and about a million times smaller than frontier models like GPT-6 Astra.
Everything shipped under the MIT licence: the model weights, the training code, and the dataset. You can download it, fork it, retrain it, and ship it inside a commercial product without asking anyone's permission. For Indian students and small teams, that last part matters more than any benchmark.
02.How a form-filling AI model works
Here's the neat engineering idea. When an AI agent fills a form today, it typically calls a big language model for every single field: "here's the document, here's the input box — what should I type?" Multiply that by 40 fields and 500 forms a day, and you're burning tokens like Diwali crackers.
CUA-S1-FORMS narrows the job to one bounded decision: given information already extracted from a document and one element on a form, decide whether to fill the field with one of the supplied values, tick a checkbox, click a button, or skip it. Crucially, it doesn't generate text at all — it scores the permitted options in a single forward pass and returns a probability for each. Cua's Driver tool then translates those decisions into actual interface actions, ordering the work so fields get filled first, checkboxes follow, and the submit button is only clicked when authorised. That last detail — submit stays gated — is the difference between automation you can trust and automation you have to babysit.
03.Why not just use GPT-6 Astra or Claude?
Honestly, for messy, judgment-heavy work, you absolutely should — big models are phenomenal at understanding ambiguous documents. But most office automation isn't ambiguous. It's the same 30 fields, every single day. Using a frontier model for that is like hiring a PhD to stamp envelopes: it works, until you see the bill at month-end. A tiny specialist that does one thing cheaply, fast and consistently is often the better engineering answer — and you can audit exactly what it's allowed to do. If you're building agents seriously, knowing when to use a big model and when to use a small one is the single biggest cost lever. For the hands-on side — actually wiring models like this into agents that run a real business workflow — our AI Agents & Automation Playbook is the practical companion. It's also a core theme in our GPT-6 Astra — Zero to Hero Guide, whose cost-optimization chapter pays for itself in your first week of API bills.
04.The bigger shift: System One models are a category now
What surprised me most is the timing. CUA-S1-FORMS landed just four days after TypeSafe AI introduced Jev, the much-talked-about model from a former OpenAI researcher that also returns typed decisions and probabilities instead of strings — we covered Jev's launch in detail earlier. Cua openly describes its model as "Jev-like", using the same broad contract of scoring a predefined set of options. When two independent teams converge on the same idea in the same week, you're watching a category being born: small, cheap, ultra-reliable decision models that sit underneath big reasoning models. It's a genuinely different architecture from the giant computer-use models we usually write about, like the Hunmin VLM release, which we covered here.
05.The India angle: we are the country of forms
Let's be real for a second. India runs on forms — GST filings, KYC onboarding, insurance proposals, college admissions, PF withdrawals, loan applications, trade licence renewals. An enormous amount of back-office employment here is humans copying data from one document into one portal, all day. A free, 2.8 MB model that can plan form-filling could power genuinely affordable automation for small CA firms, DSA networks, clinics and one-person startups — the segment that can't justify enterprise AI contracts. And because it's open weights, an Indian team can fine-tune it on Indian form formats (Aadhaar-style layouts, regional-language labels) without any vendor's blessing.
06.What to watch next
This is explicitly a family, not a one-off. Cua — whose main open-source repository has around 23,300 GitHub stars — is asking businesses with repetitive, bounded computer workflows to suggest the next specialist for the CUA-S1 line. Expect small open models for invoice matching, email triage, or data-entry validation in the coming months. If your business has a boring, high-volume task, this is the moment to start experimenting, because the tooling cost has essentially dropped to zero.
Key Takeaways
- CUA-S1-FORMS is a 706,048-parameter open-source model that plans form-filling for AI agents in a single pass
- It returns probabilities over a fixed set of actions instead of generating text — faster, cheaper, more predictable
- Model, training code and dataset are MIT-licensed; the checkpoint is just 2.8 MB
- Cua Driver executes the plan with a safety ordering: fields first, checkboxes next, submit only when authorised
- It joins TypeSafe's Jev in the emerging "System One" category of small, specialised decision models
Frequently Asked Questions
What is CUA-S1-FORMS?
An open-source model from San Francisco startup Cua with just 706,048 parameters and a 2.8 MB checkpoint, released on September 18. Its entire job is deciding how an AI agent should fill out forms. The model weights, training code and dataset are all MIT-licensed, so you can fork it and ship it inside a commercial product without asking permission.
How does a form-filling AI model work?
Instead of generating text, it scores a fixed set of permitted actions in a single forward pass and returns a probability for each — fill the field with one of the supplied values, tick a checkbox, click a button, or skip it. Cua's Driver tool then translates those decisions into interface actions, and the submit button is only clicked when authorised.
Why not just use a frontier model to fill forms?
For messy, judgment-heavy work a big model is the right tool, but most office automation is the same 30 fields every single day — using a frontier model for that is like hiring a PhD to stamp envelopes. A tiny specialist that does one thing cheaply, fast and consistently is often the better engineering answer, and CUA-S1-FORMS joins TypeSafe's Jev in the emerging 'System One' category of small, specialised decision models.
What This Means for You
If you're a developer or a final-year student, download it this weekend — a model this small runs on anything, and reading its training code will teach you more about applied AI than a month of headlines. If you run a business that handles paperwork — loans, insurance, compliance, admissions — start mapping which of your workflows are "bounded" enough for a specialist like this, because your competitors are. And if you're new to building with agents altogether, start with the fundamentals: our GPT-6 Astra — Zero to Hero Guide covers prompting, computer use, agents, cost optimization and safety in one practical package — ₹149 for the Digital edition, or ₹1,999 for Digital plus a printed book delivered anywhere in India. And if you run a CA firm or handle compliance paperwork, our CA AI Practice Playbook gives you 60 ready prompts for the GST 2.0 close, notice replies and tax audit — the kind of bounded, repeatable work a small model like this is perfect for.
One more thing — if you are still getting up to speed on GPT-6 Astra itself, our complete GPT-6 Astra walkthrough explains the model, the pricing and the privacy basics in one sitting.