Cloudflare Releases Clairic as Open Weights
Cloudflare released Clairic (27B) and Clairic-flash (9B) extraction models under Apache 2.0 on Nov. 18, with PDF input, a new architecture and $0.10 promotional pricing until Dec. 1.
Cloudflare on Wednesday, Nov. 18, released the weights of two extraction models, an updated 27-billion-parameter Clairic and a new 9-billion-parameter Clairic-flash, under the Apache 2.0 license on Hugging Face. The release takes Clairic out of the preview it entered on Oct. 19 and comes from a group the company now calls its Workers AI superintelligence team.
Key takeaways
- Clairic (27B, built on Qwen3.8-27B) and Clairic-flash (9B, built on Qwen3.5-9B) are both downloadable under Apache 2.0 and hosted on Workers AI.
- Clairic is listed at $0.24 per million input tokens and Clairic-flash at $0.10, but both cost $0.10 until Dec. 1, 2026. Output is not billed.
- On Cloudflare's own synthetic receipt test, both new models scored 576 of 576 fields, up from 369 for the preview version.

What is Cloudflare Clairic?
Clairic is an extraction model. A developer sends a block of text, images or now PDFs along with a schema of typed fields, such as a total, a date or a list of line items, and the model returns one typed value per field instead of free-form prose. Each answer carries a confidence score and a null probability, and text answers point back to the character span they came from.
"Cloudflare has become a neolab," the company wrote in its announcement, credited to Alex Reneau, Michelle Chen and Kevin Flansburg. The same team trained the Clef decision models released on Oct. 1, and Cloudflare said it is hiring for the group.
How the new Clairic differs from the preview
The October preview answered every field in a single forward pass and generated no tokens. The new version switches ordinary text and number fields to what Cloudflare calls parallel slot autoregressive generation: each field writes its own tokens and stops at its own end marker, while all fields advance together in one serving batch. Separate trained heads handle estimates, forecasts, durations, calculated dates and arithmetic.
Cloudflare says the trade-off is a small latency cost for higher accuracy. Requests can include up to four images or PDF pages in total; the model card caps embedded inputs at 4 MiB each and 8 MiB combined.
| Test | Preview | Clairic | Clairic-flash |
|---|---|---|---|
| Fresh text fields (96 requests) | 295/336 | 335/336 | 328/336 |
| Synthetic receipts, text and images | 369/576 | 576/576 | 576/576 |
| Tracked customer assertions | 1,033/1,263 | 1,263/1,263 | 1,263/1,263 |
Cloudflare ran the same internal tests on Liquid AI's D1.2, paired with Apple Vision for character recognition, which scored 304 of 336 on text fields, and on ScaleDown's Extract, which scored 259. The company called the figures "directional guidance" because no public extraction benchmark exists yet.
Latency and pricing
On a self-hosted server, Clairic answered text-field requests in 358 milliseconds at the median and 736 ms at the 95th percentile; Clairic-flash took 155 ms and 328 ms. For image fields, Clairic's median was 453 ms and Clairic-flash's 305 ms.
The 9B model that powered the preview moves to the @cf/cloudflare/clairic-flash endpoint, while the 27B model takes over @cf/cloudflare/clairic. Existing preview users are upgraded in place. Cloudflare said it absorbs the cost of output tokens and bills input only.
Why extraction models matter
Agent pipelines often call a large language model just to pull a number or date out of an email, then parse and validate the text. Cloudflare argues that small, purpose-built models can take that step faster and cheaper. The category is new: Typesafe exited stealth in September with a decision model, and Cloudflare's model work sits within its broader push to rebuild the web for agents.
What's next
The promotional price ends Dec. 1, when Clairic returns to $0.24 per million input tokens. Cloudflare said its new team is "just getting started." All accuracy and latency figures so far come from the company itself, so independent tests on the open weights will be the first outside check.
FAQ
Is Clairic open source?
The weights are. Both Clairic and Clairic-flash are on Hugging Face under Apache 2.0, which permits commercial use and self-hosting. Cloudflare had promised open weights in October without naming a license.
Which Clairic model should developers use?
Cloudflare positions the 27B Clairic for higher accuracy and the 9B Clairic-flash for cost-sensitive work. In its tests the two were close, but Clairic-flash was roughly twice as fast on text.
Does Clairic replace an LLM?
No. Cloudflare describes it as one stage in a pipeline: extraction feeds a decision model such as Clef, and a language model writes the final reply.
References
- An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team — Cloudflare Blog (Nov. 18, 2026)
- Cloudflare/clairic — Hugging Face
- Cloudflare/clairic-flash — Hugging Face
- Workers AI Changelog — Cloudflare Docs
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.