Cloudflare has turned its document reader into open weight AI. On 18 November 2026 the company released updated Clairic and a new smaller Clairic Flash, two extraction models that pull exact fields from receipts, PDFs and images, with the full weights published on Hugging Face. Both models also run as managed endpoints on Workers AI.
The idea is simple: small specialist models for narrow jobs. Instead of asking a big chat model to read a document, Clairic returns typed values such as numbers, dates and lists, with a confidence score for each field. Cloudflare calls this System One, models that answer in structured data rather than free form text.
What happened
Clairic is a redesigned model, not a small update. Cloudflare rebuilt it after the preview release, so ordinary text and number fields are now generated together in parallel instead of in a single forward pass, with better handling of tables, lists and dates. The updated Clairic has 27 billion parameters, while Clairic Flash is a lighter 9 billion parameter version built for speed.
Documents are the big upgrade. Clairic now reads PDFs as well as images, and 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 each description, quantity and price together.
Cloudflare shared its own test numbers with a clear caveat. On tracked customer assertions, both new models scored 1,263 of 1,263 correct fields, against 1,033 for the preview, and the company says to treat the figures as directional guidance since no public extraction benchmark exists yet. On fresh text fields, Clairic answered in 358 milliseconds at the median and Clairic Flash in 155 milliseconds.
| Model | Size | Workers AI price | Median response |
|---|---|---|---|
| Clairic | 27 billion parameters | 0.24 dollars per million input tokens | 358 milliseconds |
| Clairic Flash | 9 billion parameters | 0.10 dollars per million input tokens | 155 milliseconds |
Pricing is promotional for now. Both models cost 0.10 dollars per million input tokens until 1 December 2026, when Clairic rises to its standard 0.24 dollars while Clairic Flash stays at 0.10 dollars. Only input tokens are billed, and the weights carry an Apache 2.0 license for teams that want to self host. For another open model you can run yourself, read our brief on Inkling 2 Mini on a laptop.
Why it matters
Extraction is the unglamorous work that AI has to get right. Every invoice, receipt, form and contract in business needs its fields pulled out exactly, and a chat model that paraphrases a total is useless for accounting. A small model that returns typed values with confidence scores fits that job better and costs far less per call.
Open weights change who can use it. Teams can inspect, tune and self host the models instead of sending sensitive documents to a managed endpoint, which matters for finance, health and legal paperwork. Cloudflare frames the release as part of its new Workers AI superintelligence team, which builds frontier models for customers and the open source community.
What to watch
Watch whether a public benchmark appears. Cloudflare's own numbers look strong, but the company admits there is no public extraction benchmark yet, so independent tests will decide how Clairic really compares with rivals. The Hugging Face release makes those tests easy to run.
Watch the price after 1 December. The promo makes both models cost the same until then, so the real choice between Clairic and Clairic Flash starts when the larger model costs more than twice as much per million tokens. Speed against accuracy will decide which one teams pick.
Read the full post
Cloudflare has not posted an official YouTube video about this release. The full post is on the Cloudflare blog: An updated Clairic and Clairic Flash are now open weight.
Sources
Hugging Face: Cloudflare/clairic model card
Follow Learn Daily AI on Instagram at @learningdailyai for one new AI story every day.

Comments
Post a Comment