Beating AI News Flash: Cloudflare has released a new information extraction model Clairic, along with a lightweight version Clairic-flash. The two models are based on Qwen3.8-27B and Qwen3.5-9B, respectively, and both have open-sourced model weights and inference code under the Apache 2.0 license.
Clairic previously launched a 9B preview version in October, mainly used to extract amounts, dates, order numbers, and other data from text and images, and can also perform simple calculations and predictions. This time, the main model has been upgraded to 27B, adding PDF recognition, capable of processing up to 4 pages of PDFs or images simultaneously, and can more accurately recognize complex content such as tables and lists.
Cloudflare also redesigned the information extraction method. The old version gave results in a single computation, while the new version allows multiple fields to generate answers in parallel step by step, trading a small amount of speed for higher accuracy.
In official self-testing, on synthetic receipt tests containing text and images, the old version correctly extracted 369 out of 576 fields, while the new version got all answers correct. The 9B Clairic-flash also got all answers correct in the same test, with a median latency of about 155 milliseconds for processing ordinary text fields, compared to 358 milliseconds for the 27B version.
Both models are now available on Workers AI, priced at $0.10 per million input tokens until December 1. After that, the 27B version will increase to $0.24, while the 9B version remains at $0.10, with no output token fees charged for either.
