Artificial intelligence news.
Straight from the original announcements.

Cloudflare releases open-weight Clairic models for document extraction

November 18, 2026, 23:00By the AI News Bites Editorial TeamSource: Cloudflare
Cloudflare releases open-weight Clairic models for document extraction
Image: Cloudflare

Cloudflare has released updated Clairic and Clairic-flash, two open-weight models designed to extract data from text, images and PDFs. The models are hosted on Workers AI, with their full weights available on Hugging Face.

Key points
  • Updated Clairic extracts data from text, images and PDFs.
  • Clairic-flash is a faster, compact 9-billion-parameter model.
  • Both models are open-weight on Hugging Face and available on Workers AI.
  • The promotional price for both models is $0.10 per M input tokens until December 1, 2026.

A redesigned model for extraction

Cloudflare says it redesigned Clairic after its preview release to focus more closely on unstructured data. The updated model is intended for tasks such as pulling information from receipts, support tickets, commercial leases and loan pricing documents.

The company says the new architecture uses parallel slot autoregressive generation for ordinary text and number extraction fields. Each field uses the tokens it has already generated to choose its next token, then stops at its own end marker, while all fields continue advancing together in the serving batch.

Cloudflare also trained special heads for estimates, forecasts, durations, calculated dates and explicit arithmetic. The typed API keeps confidence scores, character spans and probabilities with every request, according to the company.

Clairic-flash targets faster, smaller workloads

Clairic-flash is a lighter version of Clairic with 9 billion parameters. The larger Clairic model has 27 billion parameters. Both models can process up to four images or PDF pages in total in one request, and line items can be returned as rows that keep descriptions, quantities and prices together.

Cloudflare reports that its internal testing found improved image extraction, with errors falling from 207 to 0 on benchmarks covering text and images. Its evaluation table shows Clairic and Clairic-flash each scoring 576 out of 576 on synthetic receipt fields in text and images, while Clairic scored 335 out of 336 and Clairic-flash 328 out of 336 on fresh text fields.

The company describes these results as directional because the extraction evaluations are self-reported. In latency tests using a self-hosted server and complete HTTP responses, Clairic recorded a p50 of 358 milliseconds and a p95 of 736 milliseconds on fresh text fields. Clairic-flash recorded 155 milliseconds at p50 and 328 milliseconds at p95 for the same test.

Availability, pricing and open weights

Developers can access Clairic through @cf/cloudflare/clairic and Clairic-flash through @cf/cloudflare/clairic-flash on Workers AI. Cloudflare says the API schemas remain unchanged from the preview endpoint, while the two model releases are also available on Hugging Face with full open weights.

The 27B Clairic model will cost $0.24 per M input tokens, and the 9B Clairic-flash model will cost $0.10 per M input tokens. Cloudflare says it charges only for input tokens, despite the models generating output tokens.

Until December 1, 2026, both models will cost $0.10 per M input tokens as part of a promotional discount. Applications connected to the preview Clairic endpoint will automatically receive the updated model as Cloudflare rolls it out in place. The release also marks the reintroduction of Cloudflare's Workers AI superintelligence team, which the company says is behind Clef, Clairic and other models in development.

Related news

Hugging Face page highlights DaisyRose-27B post-training efficiency
ModelsNovember 18, 2026, 20:54

Hugging Face page highlights DaisyRose-27B post-training efficiency

Hugging Face is featuring DaisyRose-27B in a page titled “DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency.” The page is dated November 18, 2026, but the…

A CatBoost model predicts fine-tuning time far better than FLOPs
ResearchNovember 18, 2026, 19:53

A CatBoost model predicts fine-tuning time far better than FLOPs

A CatBoost model predicted the end-to-end duration of fine-tuning jobs far more accurately than conventional FLOP calculations in a Hugging Face study of 272 production…

Cloudflare previews Clairic, a System One model that extracts typed values
ModelsOctober 19, 2026, 19:48

Cloudflare previews Clairic, a System One model that extracts typed values

Cloudflare has launched a preview of Clairic, the first of a new class of System One extraction models, hosted on Workers AI as @cf/cloudflare/clairic. According to…