Skip to main content

Cloudflare releases open-weight Clairic extraction models

20 NOVEMBER 2026·3 MIN READ·1 SOURCE

Cloudflare announced two Clairic models on November 18, 2026: an updated 27-billion-parameter model and a new 9-billion-parameter Clairic-flash.

Cloudflare releases open-weight Clairic extraction models

Key takeaways · 4

  • 01

    Teams can access either model through Workers AI or download open weights from Hugging Face.

  • 02

    The models accept up to four images or rendered PDF pages per request, subject to combined and per-input size limits.

  • 03

    Cloudflare reports higher accuracy in its internal evaluations, alongside a small latency cost from the redesigned architecture.

  • 04

    The promotional input-token price is $0.10 per million tokens for both models until December 1, 2026.

Two models, multiple deployment options

Cloudflare announced the release on November 18, 2026, with an updated 27-billion-parameter Clairic model and a new 9-billion-parameter model called Clairic-flash.[1] Both are available as open weights on Hugging Face and as hosted endpoints through Workers AI.[1] Cloudflare says each model repository includes weights, typed-output components, tokenizer and processor files, and Python helpers under the Apache-2.0 license.[1] Users can also run the models locally with PyTorch and Transformers without Workers AI.[1] The release follows Cloudflare’s hosted preview of Clairic on October 19, 2026.[1] Cloudflare says preview-endpoint applications will receive the updated model as it rolls out in place, while the hosted API schemas remain unchanged.[1]

Structured extraction from mixed inputs

Cloudflare says the models turn unstructured text, images and PDFs into typed values, estimates and forecasts.[1] A request can include up to four images or rendered PDF pages in total; embedded inputs are capped at 4 MiB each and 8 MiB combined, and remote media URLs are not accepted.[1] Cloudflare says a single PDF request can extract a supplier, line items and a total even when those details appear on different pages.[1] For line items, the extracted rows preserve each item’s description, quantity and price.[1] The typed interface also provides confidence scores, null probabilities and character spans for uniquely matched text evidence.[1]

Architecture changes and trade-offs

Cloudflare says ordinary text and number extraction now uses parallel slot autoregressive generation, replacing the preview’s single-forward-pass design.[1] Fields advance together, but each uses its own previously generated tokens and stops at its own end marker.[1] Cloudflare says this redesign improved text and image extraction accuracy while adding a small latency cost.[1] Separate model heads handle estimates, forecasts, durations, calculated dates and explicit arithmetic.[1] The published fresh-text response times show a speed difference between the models: Clairic’s median and 95th-percentile times are 358 ms and 736 ms, compared with 155 ms and 328 ms for Clairic-flash.[1]

Evaluation and pricing details

In Cloudflare’s internal evaluation, Clairic got 335 of 336 fresh text fields correct, compared with 295 of 336 for the preview.[1] Both updated models got 1,263 of 1,263 tracked customer assertions correct, versus 1,033 for the preview, and 576 of 576 synthetic receipt fields correct, versus 369.[1] Cloudflare says there are no public extraction benchmarks and describes comparisons with other extraction products as directional, self-reported guidance.[1] Standard Workers AI input-token pricing is $0.24 per million tokens for Clairic and $0.10 per million for Clairic-flash; a promotion sets both at $0.10 per million until December 1, 2026.[1] Cloudflare says Workers AI does not charge for output tokens, although the redesigned architecture generates them.[1]

The release gives teams a choice between hosted access and local runs, with a smaller model that Cloudflare reports responds faster on fresh-text requests. Professionals assessing extraction workflows should weigh those options against the stated input limits, latency figures and the fact that the accuracy results are internal rather than public benchmark results.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 20 November 2026

    Cloudflare releases open-weight Clairic extraction models

Sources

AI fluency, one session a day, built for your work.