Summary:

Cloudflare’s 27B and 9B extraction models are downloadable under Apache 2.0 and hosted on Workers AI. Here’s how accuracy, speed and pricing differ.

Cloudflare, Inc. released open weights for its redesigned Clairic document-extraction model and the smaller Clairic-flash on November 18. Both are also available through Workers AI. The 27B model takes over the original Clairic model ID; the new 9B model gets its own Flash ID. For document-heavy applications, the choice comes down to field accuracy, response time and what each request costs.

Where to download or call each model

ModelWorkers AI model IDDownload
Clairic 27B@cf/cloudflare/clairicWeights and runtime files
Clairic-flash 9B@cf/cloudflare/clairic-flashWeights and runtime files

The two Hugging Face repositories list Apache 2.0 and are public without a gate. The hosted model pages still label the service a preview. Cloudflare says the request schema is unchanged from the original preview, but anyone relying on the old endpoint should revalidate output against their own documents rather than assume the model itself is unchanged.

Pricing and what counts as input

Cloudflare lists Clairic-flash at US$0.10 per million input tokens. Clairic is also US$0.10 during the launch promotion, with a US$0.24 standard rate applying from December 1, 2026. Neither charges for output tokens on Workers AI. Embedded images and rendered PDF pages count as input. If the service splits a request into separate passes, it counts the supplied text and media again on each pass. Check the usage fields before estimating what a multi-field job will cost.

Clairic and Clairic-flash: extraction scores

These are Cloudflare’s internal tests, shown as correct fields or assertions out of the set total. D1.2 uses Apple Vision OCR in the comparison. The highest score is bolded in each row, including ties.

TestPrior previewClairicClairic-flashD1.2 + OCRScaleDown
Fresh text fields295/336335/336328/336304/336259/336
Mixed-input fields207/226221/226213/226116/226156/226
Tracked customer assertions1033/12631263/12631263/1263973/1263653/1263
Synthetic receipt fields369/576576/576576/576397/576280/576

Clairic led the fresh text and mixed-input rows, while both new models tied on customer assertions and synthetic receipts. Flash was faster in Cloudflare’s separate self-hosted HTTP latency test. For the 48-request fresh image-field set, median full-response latency was 305 ms for Flash versus 453 ms for Clairic. At the 95th percentile, it was 605 ms versus 976 ms. Those are measurements from that deployment and test mix, not a hosted latency guarantee.

If you regularly process invoices or receipts, run the same difficult samples through both models and count errors on the fields your workflow depends on. For PDFs, the hosted API allows up to four pages and images combined per request. A busy multi-page document may need to be split before extraction.

Sources