DEV Community

Cover image for Cloudflare Just Open-Weighted Its Document Extraction Models — Clairic 27B and a 9B Flash
Paul Lovelace
Paul Lovelace

Posted on Originally published at toolsdig.com

Cloudflare Just Open-Weighted Its Document Extraction Models — Clairic 27B and a 9B Flash

If you've ever pointed a big language model at a pile of PDFs and watched it confidently invent a total that isn't on any page, you already know why this launch matters. Cloudflare announced on November 18 that its document extraction models — Clairic and the new Clairic-flash — are now open-weight, on Hugging Face under Apache 2.0, and hosted on Workers AI. This is the graduation of the preview Cloudflare shipped on October 19, which promised open weights later. Later is now.

The interesting part isn't just that the weights are open. It's that Cloudflare went back and rebuilt the model after the preview, and the rebuild is aimed squarely at the exact places extraction usually breaks.

What actually changed since the preview

The preview model could read a photo of a receipt. The new Clairic reads PDFs too — supplier, line items and total in one request, even when they sit on different pages. You can send up to four images or PDF pages in a single request, and line items come back as rows that keep each description, quantity and price together, instead of a bag of disconnected values.

Three things changed under the hood:

  • A redesigned architecture. For ordinary text and number fields, Clairic now uses parallel slot autoregressive generation instead of a single forward pass: each field uses the tokens it has already written to choose its next token, stops at its own end marker, and all fields advance together in the serving batch. The trade-off is a small latency cost for a real accuracy gain.
  • Special heads for computation. The model got trained heads that handle estimates, forecasts, durations, calculated dates and explicit arithmetic — the stuff that turns "extract the numbers" into "give me the answer the numbers imply."
  • Better grounding. Answers stay tied to the source material, with the typed API interface intact: confidence scores, character spans and probabilities still come back with every request, so you can check where a value came from instead of taking it on faith.

Cloudflare says image reading accuracy improved from 207 errors to zero on its internal benchmarks. That number deserves the usual caveat — see the honest limits section below — but the direction is the right one: extraction that fails silently on a scanned invoice is worse than extraction that fails loudly.

Two models, two jobs

Model Size Base Best for
Clairic 27B Qwen3.8-27B Maximum extraction accuracy on hard documents
Clairic-flash 9B Qwen3.5-9B Fast, cheap, high-volume extraction

Clairic-flash is the lighter, faster sibling: strong extraction and prediction performance in a compact 9-billion-parameter model. On Cloudflare's own latency tests — complete HTTP response from a self-hosted server, all fields included — Clairic answers fresh text fields in 358ms at p50 (736ms at p95), while Clairic-flash does the same in 155ms (328ms at p95). For loan pricing workloads, flash sits at 188ms p50. Both stay under a second, which is the whole point: this is a model you call inside a request handler, not a batch job you run overnight.

The pricing — and the December 1 deadline

Here's the part to put on your calendar:

  • Until December 1, 2026: both models cost $0.10 per million input tokens.
  • From December 1: Clairic (27B) rises to $0.24 per million input tokens. Clairic-flash stays at $0.10.
  • No output-token charge. The models do generate output tokens, but Cloudflare only charges on input — it eats the output cost.

One migration note: during the preview, the 9B model lived at the @cf/cloudflare/clairic endpoint. Now the 9B moves to @cf/cloudflare/clairic-flash and the 27B takes over @cf/cloudflare/clairic. If your app is already hooked up to the preview endpoint, you automatically get the better model — Cloudflare is rolling the new 27B in place. If you want the cheaper 9B going forward, point at the flash endpoint explicitly before December 1, when the price gap appears.

How to actually use them

On Workers AI (hosted, fastest path):

@cf/cloudflare/clairic        # 27B
@cf/cloudflare/clairic-flash  # 9B
Enter fullscreen mode Exit fullscreen mode

The API schemas are unchanged from the preview endpoint, so existing integrations should keep working. You define the fields you want — typed values, not free text — and the model returns them with confidence scores and character spans pointing back at the source.

Self-hosted (open weights):

Both models are on Hugging Face as open weights under Apache 2.0, so you can run them locally, fine-tune them, or deploy them on your own infrastructure. That matters if your documents can't leave your network — medical records, legal files, financial statements — or if you want to tune extraction for one specific document type and stop paying per token entirely.

What to build with them first:

  • Receipt and invoice ingestion — supplier, line items, totals, dates, with the numbers grounded in the document
  • Support ticket triage — pull structured values (order numbers, dates, amounts) out of free-text tickets
  • Lease and loan document processing — Cloudflare's own latency tests used commercial leases and loan pricing as eval sets, which tells you where they expect the model to earn its keep
  • Expense audits and traffic forecasting — the estimation and forecasting heads are built for exactly this

The honest limits

Let's be straight about what we don't know yet:

  • There are no public extraction benchmarks. Cloudflare says so itself. Every number in this article — 576/576 fields on the synthetic receipt test (up from 369/576 for the preview model), 335/336 fresh text fields, 1,263/1,263 tracked customer assertions — is self-reported, on Cloudflare's own evals, against competitors (Liquid AI's D1.2 pipelined with Apple Vision OCR, and ScaleDown's Extract) that Cloudflare itself ran. Treat them as directional, not proof. Independent testing is what turns a launch post into a recommendation, and it hasn't happened yet.
  • The evals are Cloudflare's home turf. The test sets (receipts, leases, loan pricing, expense audits) are the workloads Cloudflare cares about. Your documents — your layouts, your handwriting, your edge cases — may behave differently.
  • Open weights, not open training data. Apache 2.0 on the weights is genuinely permissive, but this isn't a fully open reproduction of the model. Fine-tuning is yours to do; the training recipe isn't.
  • The promo ends December 1. The 27B model at $0.10 is a trial price. Budget for $0.24 from December, or plan your high-volume workloads on flash.

None of that makes the launch uninteresting — a purpose-built 27B extraction model, open-weighted, with typed outputs and source grounding, at $0.24 per million input tokens with no output charge, is a serious option for document pipelines that currently pay frontier-model prices for a job a small model can do. It just means: pilot it on your own documents before you bet a production pipeline on it.

Why this fits a bigger pattern

Cloudflare is explicit about the strategy: these are "System One" models — smaller, faster, cheaper models purpose-built for one specific task, meant to augment expensive LLM usage rather than replace it. The announcement also introduces the team behind them, the Workers AI superintelligence team, described as the frontier AI research group behind Clef and Clairic. Cloudflare is positioning itself as a model lab now, not just a host for other people's weights — and open-weight releases are how it earns developer trust while doing it.

The practical takeaway is simple: if you pay a frontier model to read documents today, you now have a cheaper, faster, purpose-built alternative you can test this week — hosted for $0.10 per million tokens until December 1, or self-hosted for free under Apache 2.0. Run your worst documents through both Clairic and Clairic-flash, check the confidence scores and character spans against the source, and let your own error rate decide.

What to do this week

  1. Try the hosted endpoint first — @cf/cloudflare/clairic on Workers AI, same schema as the preview, no setup. Feed it your ugliest real document, not a clean demo receipt.
  2. Compare both sizes on your own data — run the same documents through Clairic and Clairic-flash and check where the 9B model's answers diverge. The $0.10 promo price makes the test nearly free until December 1.
  3. Use the typed outputs — confidence scores and character spans are the whole point. Build a rule: low confidence or missing span means human review, not silent acceptance.
  4. If your documents can't leave your network, pull the Hugging Face weights — Apache 2.0, self-host, and test locally.
  5. Decide before December 1 — that's when the 27B price more than doubles to $0.24. If your volume justifies it, great; if not, move the high-volume work to flash at $0.10.

Sources: Cloudflare, "An updated Clairic and Clairic-flash are now open-weight," November 18, 2026; Unite.AI, "Cloudflare Releases Open-Weight Clairic Models for Data Extraction." Benchmark scores, latency figures and pricing are Cloudflare's own claims; no public extraction benchmarks exist yet, and the models have not been independently verified.

Top comments (0)