Cua Open-Sources Jev-Style 706K-Param Form Model, Beats Jev API 99.7% vs 83.6%

aigclink · x · 2026-09-21

Ex-OpenAI researcher Diogo Almeida released Jev, proposing a new species: the "System One" model. The core insight: models surpassed humans at chat long ago, yet automation lags because LLMs output human-readable text that programs must parse and validate. System One models invert this — no text generation, only decisions: state in, probabilistic structured choices out, all computed in parallel rather than token by token. An AI-native if-statement.

A day later, the computer-use team Cua open-sourced a scenario-specific reproduction, cua-s1-forms, for form filling:

706K parameters, 2.8MB, 2-layer Transformer, byte-level encoding
230K synthetic examples deliberately seeded with lookalike distractors (phone vs emergency-contact phone, email vs street address)
On matched tasks vs Jev's official API: 99.7% vs 83.6% (Jev is zero-shot, so the point is that task-specialized small models beat general ones)

Takeaway: every repetitive computer task will eventually get its own tiny specialist; rule-based/coordinate-recorded RPA work (forms, CRM updates, data entry, browser ops, document routing, UI decisions) can be rewritten with this paradigm — big models think, System One acts.

Original post →

Cua Open-Sources Jev-Style 706K-Param Form Model, Beats Jev API 99.7% vs 83.6%

More from Models