How many steps does a general-purpose LLM agent take to fill one web form? Cua's reference number: 23 turns and 39.6 seconds — screenshot, think, call the driver, screenshot again. Their newly open-sourced specialized model, CUA-S1-FORMS, does the same job in a single 50-millisecond forward pass.

A 706K-Parameter "System One" Model

Cua (a YC S25 team) open-sourced CUA-S1-FORMS on September 19, the first release in its CUA-S1 family. The family is named System One Models: no chain-of-thought, no multi-turn reasoning, no cloud API call — one forward pass produces the decision. CUA-S1-FORMS has just 706k parameters, a 2.8MB checkpoint, an MIT license, weights published on Hugging Face as cua-ai/cua-s1-forms, and it runs comfortably on CPU.

It does exactly one job: form decisions. Given a set of structured fields and candidate values extracted from a document, the model predicts one of four actions for each form element — use the given value, check, click, or skip. All element decisions come back in parallel from a single forward pass; your code orders the actions, and Cua Driver executes them silently one at a time. It does not predict new values, and it does not look at screenshots.

Architecture: Counter-Intuitively Small

Third-party teardowns describe a two-layer, 128-width transformer with four attention heads that reads raw bytes instead of using a tokenizer. Training was equally light: the first iteration took less than 30 minutes on synthetic data.

Self-Reported Results and Limits

The official model card reports 99.7% accuracy on real form-filling decisions, against roughly 83.6% for a general-purpose LLM API baseline. Note these are vendor-reported numbers; independent replication has not appeared yet. The limits are equally explicit: this works only where task boundaries are clear and structure is constrained. Predicting new values or understanding screenshots — the open-ended work — stays with large models.

Industry Read: Small Models Start Carving Up the Agent Pipeline

The same week, Latent.Space's AI newsletter noted that on @ValsAI's CUA-Bench, every frontier model scores below 20% on real-time keyboard-and-mouse tasks — general agents are far from saturating computer control. Against that backdrop, CUA-S1 takes the reverse route: instead of betting on ever-larger general models, it carves the "reflex-level" subtasks out of the agent pipeline and hands them to a micro-model trained in under half an hour. Needle 3, the 8-29MB on-device tool-calling model open-sourced a day earlier, is another data point in the same narrative.

For teams building enterprise automation, the signal is concrete: not every step of an agent deserves a hundred-billion-parameter model. Sinking deterministic "value alignment" subtasks into a 2.8MB local model improves latency, cost, and privacy at the same time. As for the vendor numbers — wait for independent evaluation.