THE SIGNAL IN ONE SENTENCE
Cloudflare has opened the weights for two models built to pull structured values from unstructured documents. The larger model is an updated 27-billion-parameter Clairic. The smaller one is Clairic-flash, which Cloudflare describes as a 9-billion-parameter option for faster and cheaper work. Both are hosted on Workers AI, and Cloudflare says the full weights are available through Hugging Face. This is not another general chatbot wearing a tiny accountant hat. Clairic is designed around typed extraction. A developer supplies source material and a map of fields. The response can return names, numbers, dates, lists, computed values and estimates. For ordinary extracted text, the API can include the source span. It also returns confidence information and a null probability so the surrounding system has something better than a fluent shrug. The updated model can accept text, images and PDFs. Cloudflare says one request can include up to four images or rendered PDF pages in total. Its documentation lists a 65,536-token context window for the hosted 27B model. Cloudflare also changed the machinery. For ordinary text and number fields, the new Clairic uses parallel slot autoregressive generation. In plain language, each requested field can write its answer token by token while the fields move through the serving batch together. Cloudflare says it trained separate heads for estimates, forecasts, durations, calculated dates and arithmetic. That architecture creates an interesting middle ground between old-fashioned optical character recognition and an open-ended language model. OCR tries to turn pixels into text. Clairic tries to turn the document into named, typed values. A general language model can explain the document, but it may spend more time and money producing prose you never asked for. Purpose-built is promising. Purpose-built still needs a purpose-built test. Cloudflare reports strong results on its own evaluation sets. The updated Clairic scored 335 of 336 fresh text fields, 221 of 226 mixed-input fields, 1,263 of 1,263 tracked customer assertions and 576 of 576 synthetic receipt fields. Clairic-flash also reached 1,263 of 1,263 tracked customer assertions and 576 of 576 synthetic receipt fields in the published table. Those numbers are supplied by Cloudflare. The company says there are no public extraction benchmarks yet and describes its comparisons as directional and self-reported. That caveat is doing useful work. Document extraction is unusually easy to demonstrate and unusually annoying to operate. A pristine coffee receipt can make almost any system look clever. Real document piles contain fax shadows, handwritten corrections, pages scanned sideways, totals continued on another page, empty fields, duplicate invoice numbers, locale-specific dates, tables that cross page breaks and one very important note written in the margin by somebody named Gary. Gary is always where the benchmark goes to die. The plain signal is simple: test Clairic on the document mess you actually own. Start with a gold set. Pull a representative sample from the work already moving through your organization. Include ordinary documents, but deliberately overweight the failures that create rework, payment delays, compliance risk or angry customers. For accounts payable, that might include faded receipts, multi-page invoices, purchase-order mismatches, tax lines, freight charges, credits, handwritten approvals and duplicate submissions. For support, it might include screenshots, forwarded email chains, attachments with conflicting account numbers and forms that leave required fields blank. For leases, claims or loan files, include amendments, schedules, repeated tables and dates whose meaning depends on nearby language. Do not build the test entirely from synthetic documents. Synthetic data is useful for coverage, especially when privacy limits what can leave production. It is also suspiciously tidy. The smudge, fold, crop, copier streak and human improvisation are part of the task. Label the gold set twice. Have one person extract the fields, then have another person verify the source. Resolve disagreements before the model is scored. Record the exact source page, region or text span for every field. When a value is absent, label it absent. When the document is contradictory, label it contradictory rather than forcing a convenient answer. Now separate the scores. Field accuracy asks whether each value is correct. Schema accuracy asks whether the output has the required type and shape. Row integrity asks whether a description, quantity and price stayed together instead of borrowing a value from the next line. Provenance asks whether the returned span or page actually supports the answer. Null accuracy asks whether the system leaves a missing value empty. Document acceptance asks whether the complete record can enter the next business step without repair. That final score matters most. A model can hit 98 percent of fields and still fail the operation if the two percent includes bank details, tax totals, patient identifiers or cancellation dates. Weight fields by consequence. A merchant nickname is not the same as a payment amount. An optional note is not the same as a legal deadline. Use exact matching where exactness matters. Normalize only the differences the downstream system truly treats as equivalent. A date rendered as 2026-11-19 and November 19, 2026 may be the same value. An invoice number with a missing leading zero may not be. Tables deserve their own test. Score every row as a unit. Confirm that item descriptions remain attached to quantities, unit prices, taxes and totals. Check repeated headers, blank cells, merged cells, page breaks, subtotals and negative values. Run the same table as a clean PDF, a low-resolution scan and a phone photograph. If the model changes the row structure, a perfect-looking total can hide the wrong inventory. Test source spans as evidence, not decoration. For every accepted field, follow the span back to the original state text or rendered page. Measure exact support, partial support, wrong support and no support. A span that points to the right neighborhood but not the claimed number is not provenance. It is a map with a coffee stain over the exit. The null path needs equal attention. Cloudflare's example shows a missing purchase-order number returned as null even though the model produced an internal guess. That is the behavior an extraction system should be able to express: I found something that looks plausible, but the requested field is not established. Build an abstention set with missing fields, ambiguous labels, conflicting versions and unreadable regions. Measure false fills separately from missed values. In many financial, legal and medical workflows, inventing a value is more dangerous than routing a document to review. Confidence needs calibration. Put predictions into confidence bands, then measure how often each band is actually correct. A 0.9 confidence score should not be treated as a nine-out-of-ten promise until local data shows that relationship. Calibrate separately by document family, field type, capture method and consequence. The model may be reliable on printed invoice dates and overconfident on handwritten account numbers. Turn calibration into routing. High-confidence, source-supported, low-consequence fields may flow automatically. Medium-confidence fields may need a quick comparison. Missing, contradictory, unsupported or high-consequence fields should reach a person with the relevant page already open. The reviewer should see the source, proposed value, confidence signal and reason for escalation in one place. Do not ask a person to reread the entire document because the machine felt uneasy somewhere. Measure review time and correction time. A fast model that creates a scavenger hunt may cost more than a slower model with clean evidence. Count the seconds needed to find the source, understand the failure, fix the record and resume the workflow. Then price the real request. Cloudflare says Clairic-flash costs $0.10 per million input tokens. The hosted 27B Clairic is temporarily offered at the same rate until December 1, 2026, when its listed price becomes $0.24 per million input tokens. The billing detail is more interesting than the headline price. Cloudflare's documentation says Clairic can use multiple passes. Each text field receives its own pass in some cases, while other fields share a pass. The full state and all images are included on every pass, so their tokens can be billed again. The response reports usage.passes. That means cost depends on the document, schema and pass count, not merely the file size. Record input tokens, image tokens, schema tokens, passes, latency, retries, reviewer time and corrected fields. Calculate cost per accepted document and cost per accepted high-consequence field. Compare Clairic and Clairic-flash under the same schema and acceptance rules. Latency needs the same honesty. Cloudflare reports sub-second median and 95th-percentile responses across its listed internal sets. For fresh images, it reports a 976-millisecond p95 for Clairic and 605 milliseconds for Clairic-flash. Those are company-run tests from a self-hosted server. Your path may include uploads, PDF rendering, authentication, regional network time, queueing, storage and downstream validation. Measure end to end. The user does not live inside the model timer. The four-page limit is not a footnote either. If a document exceeds the media limit, define a deterministic splitting strategy. Preserve page order, document identity, repeated headers and cross-page references. Do not let two workers process the same page as different documents. Reassemble results with page-level provenance and a conflict rule when the same field appears more than once. Longer text can fit inside the context window, but context capacity does not solve document structure. Test truncated scans, mixed orientations and files where the relevant clause refers to an appendix. If the request does not fit, fail visibly. Silent cropping is an extraction bug wearing an infrastructure badge. Open weights expand the options. Teams can inspect the published artifacts, test alternative serving stacks, run the model inside a controlled environment and build domain-specific evaluation. That can matter when documents contain trade secrets, health information, financial records or material that cannot be sent to a hosted service. Open weights do not remove operational responsibility. Self-hosting adds hardware choices, quantization, dependency security, access control, logging, patching, model versioning and performance drift. Record the exact checkpoint, file hashes, runtime, precision, hardware and preprocessing path. Run the same gold set after every change. Hosted users need a regression test too. Cloudflare says customers already using the preview Clairic endpoint will automatically receive the updated model. An automatic improvement is still a model change. Re-run protected documents before allowing the new checkpoint to approve payments, populate records or trigger customer actions. Keep the old results long enough to compare field by field. If the new model fixes three failures and creates one new high-consequence error, the release is not simply better. Finally, preserve the document boundary. Treat files as data, not instructions. A PDF can contain text that tries to redirect an agent or request a tool action. The extraction layer should only fill the allowed schema. It should not send email, change an account, approve a claim or execute code because a document told it to. Validate types after the model responds. Apply business rules in ordinary code. Require human approval for consequential actions. Keep the original file, extracted record, source evidence, model version and corrections linked in one durable receipt. Cloudflare has released an interesting piece of specialized infrastructure. Open weights make independent testing possible. Typed outputs and source spans give teams something concrete to inspect. Now feed it the ugly folder. If Clairic can survive Gary's marginalia, the split table, the missing total and the fourth-generation photocopy while leaving a clean evidence trail, then it has earned a place in the workflow. Until then, it has earned a very good test.
01
WHAT ACTUALLY CHANGED
Cloudflare released updated open weights for the 27-billion-parameter Clairic extraction model.
Cloudflare introduced Clairic-flash as a smaller 9-billion-parameter sibling.
Both models are available as hosted Workers AI endpoints and as published weights on Hugging Face.
The updated architecture uses parallel slot autoregressive generation for ordinary text and number fields.
The models can return typed values, confidence information, probabilities and source spans through the existing API shape.
The family now accepts PDFs as well as text and images, with up to four images or rendered PDF pages per request.
Cloudflare reports strong internal extraction results and sub-second latency across the evaluation sets it published.
Cloudflare says there is no public extraction benchmark yet and labels its comparisons directional and self-reported.
The hosted 27B endpoint is discounted to $0.10 per million input tokens until December 1, 2026, then lists at $0.24.
Existing preview users on the Clairic endpoint receive the updated model in place.
02
WHY THIS MATTERS
Specialized extraction models can reduce the cost and latency of turning messy documents into structured records.
Open weights allow independent evaluation and controlled deployment, including environments where documents cannot leave local infrastructure.
Source spans and null behavior can support evidence-based review when they are locally verified.
Field-level accuracy can hide document-level failures when one consequential value is wrong.
Per-pass billing means the schema and request structure can affect cost in addition to document size.
A four-page media limit requires explicit splitting, reassembly and provenance for longer documents.
An automatic hosted-model update can change production behavior even when the endpoint stays the same.
Confidence scores only become useful routing signals after calibration on local document families.
03
WHERE IT COULD HELP
- Build a gold set from real receipts, invoices, contracts, claims, forms and support documents.
- Include damaged scans, handwriting, page breaks, merged cells, conflicting dates and missing fields.
- Label every expected value with its source page, region or text span.
- Score field accuracy, schema validity, row integrity, provenance, null accuracy and complete-document acceptance separately.
- Weight mistakes by consequence so a wrong payment amount counts more than an optional nickname.
- Measure false fills separately from missed values and route unsupported fields to review.
- Calibrate confidence by document family, field type, capture method and consequence.
- Show reviewers the source, proposed value, confidence and escalation reason together.
- Track usage passes, tokens, latency, retries, reviewer time and correction time.
- Compare Clairic and Clairic-flash under identical schemas and acceptance rules.
- Define a deterministic page-splitting and result-reassembly strategy for documents longer than four pages.
- Pin checkpoint hashes, runtime, precision, hardware and preprocessing for self-hosted deployments.
- Re-run a protected regression set after hosted or self-hosted model changes.
- Keep extraction inside an allowed schema and place consequential business actions behind human approval.
KEEP A HAND ON THE WHEEL
Cloudflare supplies the published accuracy and latency results. The company says there is no public extraction benchmark yet and describes its comparisons as directional and self-reported. Perfect scores on selected internal sets do not establish reliability on a local document population, especially damaged scans, handwriting, long tables and rare high-consequence fields. The hosted model accepts up to four images or rendered PDF pages per request. Longer documents require a tested splitting and reassembly process. Cloudflare's documentation says full state and image tokens can be included on multiple passes. Measure usage.passes and total billed input rather than pricing only the original document length. Confidence, probabilities and source spans are signals to validate. They are not independent proof that a value is correct. Existing preview users receive the updated model at the same endpoint. Treat that as a model change and run regression tests before consequential automation. Open weights can support local control, but self-hosting adds model integrity, dependency, access, monitoring, patching and versioning responsibilities. The promotional price for hosted Clairic ends December 1, 2026. Confirm current pricing before budgeting a production workload.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Calibration
Measuring and adjusting a system so its controls and readouts match how the hardware actually behaves.
OPEN GLOSSARY CARD
Grounding
Connecting an AI answer to specific outside information that can support it.
OPEN GLOSSARY CARD
Audit trail
A durable record of actions, changes, identities, and times that lets someone reconstruct what happened.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on November 19, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US