Innovation & Tech

Cloudflare opens 27B Clairic weights, launches 9B flash model

Cloudflare opened the 27B Clairic and 9B Clairic-flash weights, priced hosted use at $0.10 per million input tokens through Dec. 1, 2026, and reported sharper extraction.

By Grace Kim

5 min read

Cloudflare Releases Open-Weight Clairic Models for Data Extraction
Cloudflare Releases Open-Weight Clairic Models for Data Extractionstevepb / pixabay

What's News

  • Nov. 18, 2026: Cloudflare released open weights for 27B Clairic and 9B Clairic-flash on Hugging Face and Workers AI.
  • Standard Workers AI pricing is $0.24 per million input tokens for Clairic and $0.10 for Clairic-flash.
  • Both models cost $0.10 per million input tokens through Dec. 1, 2026, with no output-token charge.
  • Clairic scored 335/336 on fresh text fields, while Clairic-flash scored 328/336.
  • Each request accepts four images or rendered PDF pages, with embedded inputs capped at 4 MiB each and 8 MiB combined.

By Grace Kim

Cloudflare announced open weights for its updated 27-billion-parameter Clairic model and new 9-billion-parameter Clairic-flash model on November 18, 2026, making both available on Hugging Face and through hosted Workers AI endpoints.

Cloudflare built the models to turn unstructured text, images and PDFs into typed values, estimates and forecasts. Customers can pay for Cloudflare's hosted inference or run the models on their own infrastructure.

What do the open-weight releases include?

The Clairic model card identifies Qwen3.8-27B as its base. Clairic-flash uses Qwen3.5-9B. Both repositories contain weights, typed-output components, tokenizer and processor files, and Python helpers under the Apache-2.0 license.

Users can run local PyTorch and Transformers inference without Workers AI. That arrangement shifts serving costs and control toward customers operating the models themselves.

The release follows Cloudflare's October 19, 2026 hosted preview, which used an early checkpoint and said open weights would follow further work. The preview already handled extraction, computation, estimation and forecasting.

Cloudflare flagged three limitations: multi-step arithmetic, estimates without supporting evidence, and basic forecasting. The open-weight release does not provide a public benchmark showing how the models perform on those tasks.

In practical terms, Cloudflare has widened deployment choice without changing the hosted API contract.

How does the redesigned architecture work?

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.

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 redesign improved text and image extraction accuracy while adding a small latency cost.

In practical terms, Cloudflare reports more accurate field extraction for a modest speed penalty.

What can one Clairic request handle?

Inputs can be text or JSON, plus optional images and embedded PDFs. Outputs include typed values, lists, nulls, estimates and forecast paths. The interface also returns confidence scores and null probabilities.

Cloudflare attaches character spans to uniquely matched text evidence. Image-only or ambiguous evidence receives null spans rather than a potentially misleading location.

Each request accepts 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.

One PDF request can extract a supplier, line items and a total even when they appear on different pages. Line-item rows retain each description, quantity and price. The page limit confines each model call to a small document set rather than an entire repository.

In practical terms, Clairic targets bounded documents rather than high-volume archives.

How strong are the accuracy and latency results?

Cloudflare's internal evaluations reported these correct-answer counts:

  • Fresh text fields, 96 requests: Clairic-preview 295/336; Clairic 335/336; Clairic-flash 328/336.
  • Tracked customer assertions, 139 request shapes: Clairic-preview 1,033/1,263; Clairic 1,263/1,263; Clairic-flash 1,263/1,263.
  • Synthetic receipt fields in text and images: Clairic-preview 369/576; Clairic 576/576; Clairic-flash 576/576.

Cloudflare says “no public extraction benchmarks exist.” Its comparisons with Liquid AI's D1.2 paired with optical character recognition and Scaledown's Extract are directional, self-reported guidance. The results document progress against Cloudflare's preview but do not establish an independent model ranking.

For fresh text fields, Clairic recorded median and 95th-percentile response times of 358 milliseconds and 736 milliseconds. Clairic-flash posted 155 milliseconds and 328 milliseconds.

On fresh image fields, Clairic recorded 453 milliseconds at the median and 976 milliseconds at the 95th percentile. Clairic-flash recorded 305 milliseconds and 605 milliseconds.

The tests measured complete HTTP responses from a self-hosted server and included every field in each request. Cloudflare's accuracy figures remain self-reported.

In practical terms, the gains are strongest against Cloudflare's own preview, not a verified industry ranking.

What changes in endpoints and pricing?

On Workers AI, Clairic now occupies @cf/cloudflare/clairic, while Clairic-flash moves to @cf/cloudflare/clairic-flash. Applications using the preview endpoint automatically receive the updated model as it rolls out in place. The hosted API schemas remain unchanged.

Customers can direct cost-sensitive workloads to Clairic-flash. Standard Workers AI pricing is $0.24 per million input tokens for Clairic and $0.10 per million input tokens for Clairic-flash. Neither model carries an output-token charge, although the redesigned architecture generates output tokens.

Through December 1, 2026, a promotion prices both models at $0.10 per million input tokens. The discount then ends for Clairic, whose standard $0.24 rate resumes.

Cloudflare also introduced the Workers AI Superintelligence team, which it describes as the frontier research group behind Clef and Clairic. It calls its System One strategy “moving narrowly defined tasks from expensive general LLM calls to smaller, purpose-built models.”

Cloudflare says its input-only pricing preserves the original System One cost model after the move to token generation. Open weights meanwhile shift serving costs and control to users who operate the models themselves.

In practical terms, customers can trade lower hosted prices for greater control over local execution.

What remains unproven?\n

Cloudflare's internal results may not transfer directly to production documents, where inputs can differ from its tests. The model cards leave acceptance thresholds for probability scores to each application, leaving customers responsible for setting their own reliability bar.

The next test is whether Clairic's reported extraction gains hold up in production document workflows without adding the latency Cloudflare measured in its own environment.

Source: unite.ai

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,387 articles

Related articles

« Previous articleNext article »