20 November 2026Explainers with cited sources
OmnistuNews
Tech & AI

Cloudflare opens weights for its Clairic extraction models, including a new 9B Clairic-flash

Cloudflare announced on 18 November 2026 that an updated Clairic extraction model and a new lighter Clairic-flash are open-weight on Hugging Face and hosted on Workers AI. The update adds image and PDF reading, and Clairic-flash packs extraction, estimation and forecasting into a 9-billion-parameter model based on Qwen3.5-9B.

AI-assisted summary of linked sources · 20 November 2026 · How we report

Graphic: Cloudflare opens weights for its Clairic extraction models, including a new 9B Clairic-flash

Cloudflare's newly named Workers AI superintelligence team released two open-weight extraction models on 18 November 2026: an updated Clairic, redesigned since its preview, and Clairic-flash, a lighter 9-billion-parameter version. Both turn a document and a schema of typed fields into structured answers with confidence scores, and both are hosted on Workers AI with full weights on Hugging Face.

What changed in Clairic

The updated Clairic adds extraction from images and PDFs, with more reliable handling of tables, lists and dates, and better grounding of answers in the source material, per Cloudflare. For text and number fields it now uses parallel slot autoregressive generation instead of a single forward pass, and special trained heads handle estimates, forecasts, durations, calculated dates and explicit arithmetic. The typed API keeps its confidence scores, character spans and per-field probabilities.

Clairic-flash

Clairic-flash is built on Qwen3.5-9B, per its Hugging Face model card. It takes a text or JSON state plus optional images and a mapping of field names to specifications, and returns typed values, lists, nulls, estimates and forecast paths with extraction confidence and null probabilities. It runs on PyTorch and Transformers with local Python helpers. Cloudflare's own eval table shows the new models beating the preview on extraction across text, images and PDFs; those are the company's internal measurements.

Context

The releases come from the team Cloudflare now calls its Workers AI superintelligence team, the group behind Clef and Clairic, which says it is building frontier models for customers and the open-source community. Open-weight extraction models are a growing niche: they let companies pull structured data from documents on their own infrastructure instead of sending it to a general-purpose chat API.

What to keep in mind

Accuracy and speed figures are Cloudflare's own internal evals, and "open-weight" means downloadable weights - check the model cards' license terms before commercial redistribution. Extraction confidence scores are model self-reports, useful for triage but not proof of correctness.

More from Omnistu News