The Pulse
Cloudflare releases open weights for Clairic and Clairic-flash
Cloudflare has made an updated Clairic extraction model and its smaller Clairic-flash counterpart available with open weights. Both run on Workers AI, with promotional pricing for the larger model through November 2026.

Cloudflare has released open weights for an updated Clairic extraction model and a smaller version, Clairic-flash, while making both available on Workers AI. The November 18 release replaces the model previously hosted at the Clairic preview endpoint with a larger version, according to the company's post.
The updated Clairic can extract information from images and PDFs as well as text, with the company reporting improvements in handling tables, lists and dates. Clairic-flash has 9 billion parameters, while the larger Clairic has 27 billion.
A redesigned extraction model
Cloudflare redesigned Clairic after its preview, changing how it generates ordinary text and number fields. The revised architecture uses parallel slot autoregressive generation instead of a single forward pass.
Each field selects its next token using the tokens it has already produced and stops at its own end marker, while the fields advance together in the serving batch. Cloudflare says the change improved extraction accuracy for text and images at a small latency cost.
The company also trained special heads to handle estimates, forecasts, durations, calculated dates and explicit arithmetic. The typed API retains confidence scores, character spans and probabilities, and its full API schemas remain unchanged from the preview endpoint.
Documents and internal tests
Requests can include up to four images or PDF pages in total. For a PDF spanning multiple pages, a request can ask for a supplier, line items and total; line items can be returned as rows that keep descriptions, quantities and prices together.
Cloudflare reports that the updated Clairic scored 335/336 on fresh text fields across 96 requests in its internal evaluation, compared with 295/336 for the preview and 328/336 for Clairic-flash. On a set of tracked customer assertions covering 139 request shapes, both new models scored 1,263/1,263, while the preview scored 1,033/1,263.
The company says the evaluations are self-reported and should be treated as directional guidance. Its comparison also included Liquid AI's D1.2, paired with Apple Vision for optical character recognition, and ScaleDown's Extract.
Latency tests measured complete HTTP responses from a self-hosted server, including all requested fields. For 96 fresh-text requests, Cloudflare reported p50 and p95 latencies of 358 and 736 milliseconds for Clairic, and 155 and 328 milliseconds for Clairic-flash.
Endpoints and pricing
The preview endpoint previously hosted the 9B variant. The 9B model is moving to @cf/cloudflare/clairic-flash, while the 27B model will use @cf/cloudflare/clairic; applications using the preview Clairic endpoint will receive the new model as it is rolled out in place.
Until December 1, 2026, both models cost $0.10 per million input tokens. After the promotion ends, Clairic will cost $0.24 per million input tokens, while Clairic-flash will remain at $0.10 per million input tokens. Cloudflare charges only for input tokens, even though the models generate output tokens.
Source
Related stories
- Cloudflare Finds 155 GLM-5.3 Responses Exposed Other Users’ Data
- Hugging Face switches language-model routing to OpenRouter
- Reflection details Gray Swan safety tests of open-weight Beam model
- Cloudflare open-sources experimental Kitesurf browser for AI agents
- Cloudflare Releases Clef Decision Models and RL Fine-Tuning Service