Redact’s suggestion model is Laya, a 421-million-parameter decision model, in its 8-bit form: the LAYA_421M_Q8_0 constant in QVAC. It is 450 MB on disk. Choosing something this small was deliberate.

​
What the model actually has to do

The model’s job in Redact is narrow. It reads at most a page’s worth of short OCR lines and answers a fixed set of yes/no questions about each one. It doesn’t write prose, doesn’t need world knowledge, and doesn’t decide labels on its own: deterministic rules re-derive the label from the text and drop anything they can’t support (About grounded suggestions). With that much scaffolding, a small model is enough to pass Redact’s labelled evaluation, scoring 0.83 to 0.92 F1 on the four test pages. Where it falls short, a different model might help, but so might better OCR or a better-worded question; the evaluation output shows which field was missed.

​
What small buys

Size shows up everywhere on a local machine:
  • Download: 450 MB is a few minutes on a home connection. General-purpose models that can follow instructions are often several gigabytes.
  • Memory: a small model leaves room for the OCR engine, the browser and the reader’s other work.
  • Latency: Laya decides a page in under a second in testing. OCR already takes 20 to 30 seconds, so the suggestion step stays out of the way.

​
When a different model is worth it

QVAC ships other Laya checkpoints: an F16 version of the same model (LAYA_421M_F16), a typed-decisions variant (LAYA_TYPED_DECISIONS_421M_Q8_0), and a multilingual one with 322 million parameters (LAYA_MULTILINGUAL_322M_Q8_0). The multilingual model is the obvious next step for documents that aren’t in English; the OCR model is a Latin-script recognizer, though Redact currently asks it for English only. None of them has been measured on Redact’s evaluation yet. A different model won’t help with text OCR misread, with fields the grounding rules then drop, or with speed. And it changes the evaluation baseline, so the labelled gate should be re-run before trusting it (About measuring suggestions). QVAC can estimate whether a model fits a device before you download it: Assess model fit compares candidates against the machine’s memory.

​
The general point

Model size is a product decision, not a quality knob to turn up. Redact gets accuracy from the system around the model: OCR evidence, typed questions, rule-based grounding and human review. A system built that way can use a small model and stay predictable.