Cloudflare has released the weights for Clairic, a 27-billion-parameter extraction model, and Clairic-flash, a 9-billion-parameter sibling, on Hugging Face. Both also run on Workers AI, Cloudflare’s hosted inference service. The company says they are meant to take over a narrow job from expensive general LLMs: reading unstructured input and returning typed fields.

The release also rewires an endpoint that developers already call. During the preview, the 9B model sat at @cf/cloudflare/clairic. That address now serves the 27B model, and the 9B model moves to @cf/cloudflare/clairic-flash, according to Cloudflare’s blog post of November 18. Anyone who wired an app to the preview endpoint gets the larger model without changing code.

That swap carries a pricing catch. Cloudflare charges only for input tokens: $0.24 per million for Clairic and $0.10 per million for Clairic-flash. Until December 1, 2026, Clairic is discounted to $0.10 as well. After that date the same traffic on the same endpoint costs 2.4 times as much, unless the team points it at Clairic-flash.

The architecture changed from the preview. Cloudflare says ordinary text and number fields are now generated by parallel slot autoregressive decoding, where each field conditions on its own earlier output and halts on a private terminator, replacing the earlier one-shot pass. Dedicated heads cover forecasts, estimates, time spans, derived dates, and arithmetic. The typed API is unchanged, and every answer still carries a confidence score, character spans, and probabilities.

Document handling is the headline feature. One request can now carry four images or PDF pages at most, so a supplier name, line items, and a total can come back even when they sit on different pages. Line items return as rows that keep description, quantity, and price together.

The numbers come from Cloudflare’s own tests, and the company says so. On 96 fresh text-field requests (336 fields), the preview model got 295 right, Clairic got 335, and Clairic-flash got 328. Liquid AI’s D1.2, paired with Apple Vision for character recognition, scored 304, and Scaledown’s Extract scored 259. Cloudflare also reports image-reading errors falling from 207 to zero on its internal benchmark.

Treat those results as a vendor demo, not a verdict. Cloudflare concedes that no public benchmark for this task exists yet. One test set (139 request shapes of tracked customer assertions) shows Clairic and Clairic-flash both at a perfect 1,263 of 1,263, which tells you the set no longer separates the two models. The comparison models were run by Cloudflare, not by their makers.

Speed is where the small-model case is strongest. For fresh text fields, Cloudflare reports a median response of 358 milliseconds for Clairic and 155 milliseconds for Clairic-flash, measured on a self-hosted server. The slowest 5 percent of Clairic text requests finished in 736 milliseconds. Those figures cover the full HTTP response, not just generation.

The strategic point sits behind the model card. Cloudflare calls itself a neolab and has formed a Workers AI superintelligence team, which previously shipped the open-weight Clef model. Extraction is an unglamorous bet, but it is one where a company that sells inference by the token can undercut a frontier API on both price and delay.

For teams running invoice, receipt, or ticket parsing through a large model today, the useful test is cheap to run: send a few hundred of your own documents through Clairic-flash before December 1, and compare error counts against your current model before the 27B price returns.

Cloudflare, “An updated Clairic and Clairic-flash are now open-weight” (blog.cloudflare.com), published November 18, 2026.