Cloudflare releases 2 open models that read data from documents
Cloudflare says image errors fell from 207 to 0 on its internal tests, but no other test confirms this.
Cloudflare, a company that operates web infrastructure, released an updated Clairic model and a new, smaller model, Clairic-flash. The 2 models read text, images and PDF files. They give the data that you want, such as the supplier and total of an invoice. The weights are open on a site that shares models. Cloudflare says image errors fell from 207 to 0 on its internal tests.
How it works
The new Clairic writes each field one piece of text at a time, and all fields move together. Cloudflare says a response takes less than one second. One query can have a maximum of 4 images or PDF pages.
The numbers
Cloudflare says image errors fell from 207 to 0 on its internal tests. It also compared the models with other extraction models, such as the model of Scaledown. Cloudflare calls these results directional, because Cloudflare reports them.
What is not known
At this time, no test by other companies exists. Cloudflare gave no data on the test set behind the 0 errors. Other persons did not do tests on the models.
Insight: The low price is an offer for a short time
Cloudflare charges only for input and pays the cost of output. Clairic costs $0.10 for each million input tokens (small pieces of text) until 1 December 2026, then $0.24. Clairic-flash stays at $0.10. It is possible that Cloudflare wants work on its platform first. The open weights let you move to a different platform. Compare the price after 1 December before you commit.
Sources
Posted