Cloudflare opens weights for Clairic extraction models
Cloudflare has released two open-weight extraction models: an updated Clairic, which leaves preview, and a new, smaller Clairic-flash. Both run on Workers AI, and the full weights are on Hugging Face. Cloudflare presents
Cloudflare has released two open-weight extraction models: an updated Clairic, which leaves preview, and a new, smaller Clairic-flash. Both run on Workers AI, and the full weights are on Hugging Face. Cloudflare presents them as the work of its new Workers AI superintelligence team.
What changed in the models
Cloudflare redesigned Clairic after the preview. For ordinary text and number fields it now generates each field's answer token by token, with all fields advancing together in one serving batch, instead of a single forward pass. The company says this raised accuracy for text and images and costs only a little latency. Separate heads handle estimates, forecasts, durations, calculated dates and explicit arithmetic.
The API keeps confidence scores, character spans and probabilities with every response. Clairic-flash is a 9-billion-parameter version built to be lighter and faster. The 27B model sits at the Clairic endpoint.
Clairic now accepts PDFs as well as images. A single request can include up to four images or PDF pages, so a supplier name, line items and a total can come back together even when they sit on different pages. Line items return as rows that keep description, quantity and price together.
Numbers, with a caveat
Cloudflare's own evals show Clairic scoring 1,263 of 1,263 on tracked customer assertions, against 1,033 for the preview. Clairic-flash also scored 1,263. On synthetic receipt fields in text and images, both new models scored 576 of 576, against 369 for the preview. Cloudflare also compared Liquid AI's D1.2 paired with Apple Vision OCR and Scaledown's Extract, which scored lower on its tests. Cloudflare itself says the figures are self-reported and directional, and that no public extraction benchmarks exist yet.
On latency, measured as full HTTP responses from a self-hosted server, Clairic took 358 ms at the median and 736 ms at p95 on fresh text fields. Clairic-flash took 155 ms and 328 ms. Image fields were slower: 453 / 976 ms for Clairic and 305 / 605 ms for Clairic-flash.
Pricing and migration
The preview endpoint ran the 9B model at $0.10 per million input tokens. The 9B model now lives at @cf/cloudflare/clairic-flash, and the 27B model at @cf/cloudflare/clairic. Only input tokens are billed; Cloudflare absorbs the output cost.
Standard pricing is $0.24 per million input tokens for Clairic and $0.10 for Clairic-flash. Until December 1, 2026, Clairic is also $0.10. On that date it goes to $0.24.
Apps already pointing at the Clairic endpoint get the new model in place. The API schemas are unchanged from the preview, according to Cloudflare, which encourages cost-sensitive workloads to move to Clairic-flash.
Open weights
Publishing the weights on Hugging Face lets teams inspect, tune and self-host the models. Cloudflare says the aim is to let companies pair small task-specific models with their more expensive LLM usage.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.