# Cloudflare makes extraction models Clairic and Clairic-flash open-weight

By Ivo Donker · 19 November 2026 — text: Claude (AI)
What is Cloudflare releasing? Two models that pull data out of documents, Clairic (27 billion parameters) and the smaller Clairic-flash (9 billion), with open weights on Hugging Face. They read text, images and now PDFs, and return fields such as supplier, line items and total. On Workers AI both cost 0.10 dollars per million input tokens until December 1. After that Clairic moves to 0.24 dollars; output tokens are free.

We read [Cloudflare's blog post](https://blog.cloudflare.com/open-clairic) of November 18, written by Alex Reneau, Michelle Chen and Kevin Flansburg. The topic was also covered in the [trending briefing of November 19](https://trending.llmnet.nl/en/briefing/2026-11-19).

## What happened

Clairic first appeared as a preview on Workers AI. That preview endpoint ran the small 9B variant, because it was faster and easier to train. That model now moves to @cf/cloudflare/clairic-flash, and the larger 27B model runs at @cf/cloudflare/clairic. Anyone already connected to the preview endpoint gets the new model automatically.

Cloudflare changed the design. The preview filled every field in a single pass; the new Clairic generates ordinary text and number fields in parallel, step by step. Separate components were trained for arithmetic, estimates, forecasts and calculated dates. The API stays the same: every answer still carries confidence scores, the position in the source text and probabilities.

PDFs are new. You can send up to four images or PDF pages per request, and invoice lines come back as rows that keep description, quantity and price together. According to Cloudflare, tables, lists and dates are handled more reliably.

## Prices and benchmarks

Prices and benchmarks
Test (correct fields) | Preview | Clairic | Clairic-flash | 

Fresh text fields, 96 requests | 295/336 | 335/336 | 328/336 | 

Internal test, mixed input, 36 requests | 207/226 | 221/226 | 213/226 | 

Customer assertions, 139 request shapes | 1,033/1,263 | 1,263/1,263 | 1,263/1,263 | 

Synthetic receipts in text and images | 369/576 | 576/576 | 576/576 | 

Price per million input tokens after December 1 | - | 0.24 dollars | 0.10 dollars | 

Cloudflare also compared against Liquid AI's D1.2 (paired with Apple Vision for text recognition) and Scaledown's Extract; both scored lower on every test. There are no public extraction benchmarks yet, and Cloudflare itself calls its figures directional. On a self-hosted server a complete answer usually arrived within a second; Clairic-flash was one and a half to more than two times as fast as Clairic.

## What this means

Document extraction (receipts, invoices, support tickets) is one of the most practical uses of language models, and this is a specialised model with open weights. The 9B variant is small enough to try locally or on your own GPU. The blog does not name a license, but both repositories on Hugging Face are listed under Apache 2.0. According to that metadata Clairic is derived from Qwen3.8-27B and Clairic-flash from Qwen3.5-9B.

If you already use the preview endpoint, watch the price. You get the larger model automatically, and from December 1 it costs 0.24 instead of 0.10 dollars per million input tokens. If cost matters more than the last few percent, switch to clairic-flash. We wrote earlier about [the KV cache leak in Workers AI](https://nieuws.llmnet.nl/en/cloudflare-workers-ai-kv-cache-lek). Other recent open-weight models are [Inkling 2 Small](https://nieuws.llmnet.nl/en/inkling-2-small-gewichten) and [the multilingual Lara Mini](https://nieuws.llmnet.nl/en/lara-mini-meertalig-basismodel).

## What we did not verify

The license and base model come from Hugging Face's model metadata (queried on November 19); we did not read the full model cards or the documentation on developers.cloudflare.com. We did not test whether the 9B variant runs at a usable speed on a Mac or without an Nvidia card. The test results come from Cloudflare and were not re-run; the latencies were measured on a self-hosted server, not on Workers AI.

Checked on November 19, 2026.

This article was written with AI (Claude) and checked against the sources below.

## Sources

- Cloudflare, An updated Clairic and Clairic-flash are now open-weight (November 18, 2026): [blog.cloudflare.com](https://blog.cloudflare.com/open-clairic)

- Clairic weights on Hugging Face: [huggingface.co](https://huggingface.co/Cloudflare/clairic)

- Clairic-flash weights on Hugging Face: [huggingface.co](https://huggingface.co/Cloudflare/clairic-flash)

- @corywilkerson on X about Clairic: [x.com](https://x.com/corywilkerson/status/2123182780070949275)
