AI Models & Platforms
Cloudflare Releases Open-Weight Clairic Models for Data Extraction

Key PointsOpen weights shift serving costs and control from Cloudflare to users running Clairic…
View more
- Open weights shift serving costs and control from Cloudflare to users running Clairic locally.
- Parallel slot autoregressive generation improved extraction accuracy at a small latency cost.
- Cloudflare's accuracy figures are self-reported and lack independent public extraction benchmarks.
Cloudflare announced on November 18, 2026 that its updated 27-billion-parameter Clairic and new 9-billion-parameter Clairic-flash are available with open weights on Hugging Face and as hosted Workers AI endpoints. Both are built to turn unstructured text, images and PDFs into typed values, estimates and forecasts.
The Clairic model card identifies Qwen3.8-27B as its base, while the Clairic-flash model card uses Qwen3.5-9B. Both repositories include weights, typed-output components, tokenizer and processor files, and Python helpers under the Apache-2.0 license, allowing local PyTorch and Transformers inference rather than requiring Workers AI.
The release follows Cloudflare’s October 19, 2026 hosted preview, which used an early checkpoint and said open weights would follow further work. That preview already covered extraction, computation, estimation and forecasting, while its guidance flagged multi-step arithmetic, estimates without supporting evidence and basic forecasting as limitations.
A Redesigned Extraction Architecture
For ordinary text and number extraction, Cloudflare replaced the preview’s single-forward-pass design with parallel slot autoregressive generation. Fields advance together in the serving batch, but each field uses its own previously generated tokens to select the next token and stops at its own end marker. Separate heads handle estimates, forecasts, durations, calculated dates and explicit arithmetic. Cloudflare reports that the change improved text and image extraction accuracy while adding a small latency cost.
The typed interface retains confidence scores, null probabilities and character spans for uniquely matched text evidence; image-only or ambiguous evidence has null spans. Inputs can be text or JSON, optional images and embedded PDFs, while outputs include typed values, lists, nulls, estimates and forecast paths. Requests are limited to four images or rendered PDF pages in total, with embedded inputs capped at 4 MiB each and 8 MiB combined, and remote media URLs are not accepted.
PDF support allows one request to ask for a supplier, line items and total even when they appear on different pages, with line-item rows preserving each description, quantity and price. The page cap defines the scope of each model call: it handles a small set of pages rather than an entire document repository.
Self-Reported Accuracy and Latency
Cloudflare’s internal evaluations report these correct-answer counts for the updated models and the earlier preview checkpoint:
| Evaluation | Clairic-preview | Clairic | Clairic-flash |
|---|---|---|---|
| Fresh text fields, 96 requests | 295/336 | 335/336 | 328/336 |
| Tracked customer assertions, 139 request shapes | 1,033/1,263 | 1,263/1,263 | 1,263/1,263 |
| Synthetic receipt fields in text and images | 369/576 | 576/576 | 576/576 |
The company says no public extraction benchmarks exist and describes its comparisons with Liquid AI’s D1.2 paired with optical character recognition and Scaledown’s Extract as directional, self-reported guidance. The figures therefore document Cloudflare’s reported progress against its preview; they do not establish an independent ranking of extraction models.
Cloudflare reports median and 95th-percentile response times of 358 ms and 736 ms for Clairic on fresh text fields, compared with 155 ms and 328 ms for Clairic-flash. On fresh image fields, Clairic recorded 453 ms and 976 ms at those same percentiles, while Clairic-flash recorded 305 ms and 605 ms. The tests measured complete HTTP responses from a self-hosted server and included every field in each request.
Endpoint Migration and Input-Token Pricing
On Workers AI, the larger model now occupies @cf/cloudflare/clairic, while the smaller variant moves to @cf/cloudflare/clairic-flash. Applications using the preview Clairic endpoint automatically receive the updated model as it rolls out in place, and the hosted API schemas remain unchanged. Customers can instead point cost-sensitive workloads to the cheaper Clairic-flash endpoint.
Cloudflare set standard Workers AI pricing at $0.24 per million input tokens for Clairic and $0.10 per million input tokens for Clairic-flash, with no output-token charge although the redesigned architecture does generate output tokens. Until December 1, 2026, a promotion prices both models at $0.10 per million input tokens. The discount then ends for Clairic and its standard rate resumes.
The releases also introduce the Workers AI Superintelligence team, which Cloudflare describes as the frontier research group behind Clef and Clairic, with a System One strategy of moving narrowly defined tasks from expensive general LLM calls to smaller, purpose-built models.
Cloudflare’s input-only pricing preserves the original System One cost model even after the move to token generation, while open weights shift serving costs and control to users who run the models themselves. The open question is how the internal results transfer to production documents, where the model cards leave acceptance thresholds for probability scores to each application.












