Cloudflare Opens Weights for Clairic Extraction AI

Cloudflare has released open weights for its document-reading AI models, turning a preview tool into a pair of small, specialised models that anyone can download and run. The company announced on 18 November that an updated version of Clairic is leaving preview, joined by a lighter, faster version called Clairic-flash, with both hosted on Workers AI and their full weights published on Hugging Face.
Clairic is built for one job: pulling clean, typed fields out of messy inputs such as receipts, support tickets and documents. Instead of returning free-form text, it returns structured values such as numbers, dates and categories, with confidence scores, character spans and probabilities attached to each answer. Cloudflare calls these "System One" models, and says their purpose is to let companies pair small, task-specific models with more expensive large language models rather than using a frontier model for every job.
Rebuilt after the preview
Cloudflare says customer feedback after the preview pushed it back to the drawing board. For ordinary text and number fields, the redesigned model now uses what the company calls parallel slot autoregressive generation instead of a single forward pass. Each field uses the tokens it has already written to choose its next token, then stops at its own end marker, while all the fields advance together in one serving batch.
The update also adds PDF reading. In the preview, Clairic read a photo of a receipt. Now the Clairic family can read PDFs too, so a single request can ask for the supplier, line items and total even when they appear on different pages. A request can include up to four images or PDF pages in total, and line items can come back as rows that keep each description, quantity and price together. Handling of tables, lists and dates has also been improved, with better grounding of answers in the source material.
The flagship Clairic model has 27 billion parameters, while Clairic-flash has 9 billion. The 9 billion model previously lived at the preview Clairic endpoint; it now moves to the Clairic-flash endpoint, and the 27 billion model takes over the main Clairic name. Applications already calling the preview endpoint will automatically receive the new model as it rolls out in place.
What Cloudflare's own tests show
Cloudflare's internal evaluations show a sharp improvement over the preview, but the company itself warns that the figures are self-reported and should be treated as directional guidance, since there are no public extraction benchmarks yet. On tracked customer assertions, both new models scored 1,263 out of 1,263, against 1,033 for the preview. On fresh text fields across 96 requests, Clairic scored 335 out of 336, against 295 for the preview, and Clairic-flash scored 328. On synthetic receipt fields in text and images, both new models scored 576 out of 576, against 369 for the preview.
Latency is measured as a complete HTTP response from a self-hosted server. Clairic took 358 milliseconds at the median and 736 milliseconds at the 95th percentile on fresh text fields, while Clairic-flash took 155 and 328 milliseconds. Image fields were slower: 453 and 976 milliseconds for Clairic, and 305 and 605 for Clairic-flash. Even with the redesigned architecture, Cloudflare says responses arrive in under one second, much faster than typical large language model responses for similar work.
Pricing follows the same pattern. The 27 billion Clairic model will cost 0.24 dollars per million input tokens, while the 9 billion Clairic-flash stays at 0.10 dollars. Until 1 December 2026, both cost 0.10 dollars per million input tokens under a promotional discount. Cloudflare charges only for input tokens and absorbs the cost of generating output tokens. The company also introduced itself in the announcement as a "neolab" building frontier models for customers and the open-source community, with a new Workers AI superintelligence team behind Clairic and its earlier model Clef.