Open-weight extraction models, multimodal documents and structured inference

Cloudflare opens Clairic and Clairic-flash weights for structured document extraction

An updated 27B extraction model and a smaller 9B variant bring typed answers, PDF reading and source-grounding to downloadable checkpoints, with benchmarks still self-reported.

Open Model Weights published20 Nov 2026
Primary sourceCloudflare — official Clairic and Clairic-flash open-weight release
Source published2026-11-18
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original minimalist editorial illustration representing vision in AI research; conceptual, not actual model output.
AI-generated editorial illustration · Not official model imagery.

Cloudflare has moved its Clairic extraction family from a hosted preview into the open-weight ecosystem. On November 18, the company announced downloadable weights for a redesigned Clairic model, alongside a new, smaller Clairic-flash. Cloudflare describes the models as roughly 27 billion and 9 billion parameters respectively. Both are available as hosted Workers AI endpoints, while their official Hugging Face repositories provide a route for teams that want to inspect and deploy the underlying model artifacts themselves. Primary source ↗

The distinction from a conventional chatbot is the output contract. Rather than asking a general-purpose language model to write a free-form answer and then parsing it, Clairic is built to return specified typed fields from unstructured inputs. Cloudflare's examples include receipts, commercial documents, support material, numerical estimates and calculated dates. Responses can retain confidence scores, probability information and character spans, giving developers more structure for downstream verification.

The architecture has also changed since the preview. Cloudflare says its updated model generates each extraction field autoregressively within a parallel batch, rather than relying on a single pass for every value. That change aims to improve fidelity on text-heavy and image-based documents, while preserving the typed API. Newly supported PDF reading allows a request to extract related fields spread across pages. The company's current stated input limit is a combined total of four images or PDF pages.

Cloudflare reports improved internal extraction accuracy and publishes latency measurements for both variants. For fresh text fields, its table lists median full HTTP response times of 358 milliseconds for Clairic and 155 milliseconds for Clairic-flash. Those figures come from the publisher's own testing. They are not hardware-neutral benchmarks, and the release explicitly notes that public extraction benchmarks are not yet available. OMW has not reproduced the measurements.

The open-weight release is more consequential than Cloudflare's temporary hosted pricing. Teams can now evaluate whether a specialized extraction model reduces the cost and complexity of an information-processing pipeline, and can compare the two checkpoints under their own document distributions. Reproducible adoption still requires reviewing the exact weight revisions, licensing files, preprocessing components and local runtime before treating a Hub listing as a verified deployment.

For enterprise users, the key question is not whether a receipt demo looks convincing. It is whether extraction remains accurate on damaged scans, complex tables, multilingual documents and fields requiring arithmetic, and whether confidence scores are well calibrated. Those are testable questions now that the weights are public. Until independent evaluation is available, vendor-reported improvements should remain clearly labeled as such.

OMW REGISTRY WATCH

What this changes for the evidence layer

Announced full downloadable open weights on official Cloudflare Hugging Face repositories. OMW has not completed exact-SHA, file-by-file safetensors, license-file and runtime audits; do not add either to the verified model registry or label a license from the announcement alone. Cloudflare's extraction and latency metrics are internal.

PRIMARY SOURCE

Cloudflare — official Clairic and Clairic-flash open-weight release

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗