Cloudflare has released updated Clairic extraction models that can turn text, images and PDFs into structured fields in under a second in the company’s tests. The 27-billion-parameter Clairic and 9-billion-parameter Clairic-flash are available through Workers AI, with full weights published on Hugging Face.
The models preserve useful audit details alongside each answer, including confidence scores, probabilities and character spans pointing back to the source. They can extract rows that keep descriptions, quantities and prices together across as many as four images or PDF pages. Special prediction heads also handle calculations, durations and estimated dates.
Cloudflare reports that the larger model answered 335 of 336 fresh text fields correctly and both versions completed every tracked customer assertion in an internal evaluation. The company explicitly describes those comparisons as directional because there is no established public extraction benchmark and the results are self-reported.
Workers AI charges only for input tokens. Clairic-flash costs $0.10 per million input tokens, while Clairic will cost $0.24 after a temporary discount ends December 1. Self-hosting the open weights gives teams another route, but they will need to reproduce accuracy and latency on their own documents.