Cloudflare open-weights Clairic and Clairic-flash: the 27B and 9B extraction models move to parallel slot generation, and internal image-reading errors fall to zero
Synopsis
Cloudflare, under its Workers AI superintelligence team, released an updated Clairic and a lighter Clairic-flash, both open-weight and hosted on Workers AI. The update replaces a single forward pass for ordinary text and number fields with parallel slot autoregressive generation, where each field picks its next token from the tokens it has already written and stops at its own end marker, and adds special heads for estimates, forecasts, durations and calculated dates. In Cloudflare's self-reported internal evals, mixed-input fields rise from 207/226 in the preview to 221/226, and image-reading errors fall from 207 to 0; there are no public extraction benchmarks, and the authors call these evals directional guidance.
Interpretation
The updated Clairic extends extraction from photos to PDFs, accepting up to four images or PDF pages in one request so supplier, line items and total can be retrieved across pages, with line items returned as rows that keep each description, quantity and price together. The preview Clairic could read a photo of a receipt but not PDFs. In Cloudflare's self-reported internal evals, image-reading errors fall from 207 to 0, explicitly covering text and images.
Extraction of ordinary text and number fields no longer relies on a single forward pass; each field uses the tokens it has already written to choose its next token and stops at its own end marker, with all fields advancing together in the serving batch. The preview used a single forward pass for these fields, without step-by-step generation per field. On 96 requests of fresh text fields, Clairic scores 335/336 against 295/336 for the preview, 304/336 for D1.2 plus OCR and 259/336 for ScaleDown.
The models add special heads for estimates, forecasts, durations, calculated dates and explicit arithmetic, while the typed API keeps confidence scores, character spans and probabilities. The preview had no such computation and estimation heads. On 139 request shapes of tracked customer assertions, Clairic and Clairic-flash both score 1,263/1,263, against 1,033/1,263 for the preview.
Clairic-flash is the compact 9-billion-parameter variant and Clairic the 27B one, served at @cf/cloudflare/clairic-flash and @cf/cloudflare/clairic respectively, with open weights published on Hugging Face. The preview endpoint @cf/cloudflare/clairic previously hosted the 9B variant. On synthetic receipt fields in text and images, Clairic and Clairic-flash both score 576/576, against 369/576 for the preview.
Perspective
The demonstrated scope is Cloudflare's self-reported internal extraction evals and HTTP response latency tests on a self-hosted server, covering request sets such as fresh text fields, fresh image fields, commercial leases, loan pricing, traffic forecasting and expense audit; the latency figures measure the complete HTTP response including all fields in the request.
Cloudflare states there are no public extraction benchmarks yet and that the listed evals are self-reported and should be taken as directional guidance; the latency tests measure complete HTTP responses from a self-hosted server rather than end-to-end behavior of the hosted service.
