Cloudflare is best known for running a large slice of the internet's plumbing. This week it also became a model maker: on November 18, its new Workers AI Superintelligence team released Clairic and Clairic-flash as open weights, two models built to do exactly one job, pulling structured data out of documents.

Feed Clairic a receipt, a support ticket, an image or a PDF, and it does not write you a paragraph about it. It returns named, typed fields: supplier, line items, totals, and each field comes with a confidence score, so the code downstream can accept what the model is sure about and route the rest to a human. That is the whole product: documents in, data your software can use out.

There are two sizes. Clairic is the full model, at 27 billion parameters. Clairic-flash is a 9-billion-parameter version that, by Cloudflare's own figures, answers in well under a second. Both are free to download from Hugging Face under the Apache-2.0 licence, use them, modify them, run them commercially, and both are also hosted on Workers AI, if you would rather not run a model yourself.

One honest caveat: there are no public benchmarks for document extraction yet, and Cloudflare says so itself. Treat any quality claim, including theirs, as unproven until you have run the models on your own documents, with your own messy scans and odd layouts.

Why it matters

Extraction is the front door of almost every real document pipeline, and of most retrieval systems: before you can search, summarise or act on a document, something has to turn it into fields. Today that something is often the biggest, most expensive model a team has, quietly burning frontier-model money on invoices and forms. A small, purpose-built model you can run yourself changes that math: cheaper per page, your data never leaves your own environment, and the confidence score gives you a natural review queue instead of silent errors.

It is also a pattern worth learning in its own right: match the model to the job instead of defaulting to the flagship. That judgement, what to build, what to buy, and what a smaller tool can handle, is most of applied AI engineering.

Sources: Cloudflare's announcement and Unite.AI's coverage.

Go deeper

Extraction is where every RAG system in the course begins. The full interactive course, 11 chapters free, no signup, is at deployedlab.dev.

← All the latest in AI