Cloudflare has released the weights of its extraction models. On the evening of November 18, 2026, the company published a reworked version of Cloudflare Clairic with 27 billion parameters along with the smaller Clairic-flash at nine billion. Both read structured values from text, images and now PDFs, and both are available on Hugging Face under the Apache 2.0 license. At the same time, the team behind them confidently introduced itself as the “Workers AI superintelligence team”. For users of the earlier preview, however, the pricing is changing too.
Cloudflare Clairic: what the models do
Clairic is not a chat model. Cloudflare places it in a class it calls “System One”, borrowing Daniel Kahneman’s distinction between fast, intuitive thinking and slow, deliberate thinking. The idea is small, specialized models for recurring tasks that companies often hand to expensive large language models today: What is the invoice total? When was it delivered? How many days overdue is a payment?
According to the announcement on the Cloudflare blog, Clairic takes a state, meaning text or JSON plus optional images, and a schema of typed fields. It returns typed values, lists, nulls, estimates and forecasts. For every field, the model reports confidence scores, character spans in the source and the probability that a value is missing. Cloudflare trained dedicated heads for arithmetic, estimates, durations and calculated dates.
Cloudflare first presented Clairic as a preview on October 19 and promised open weights “soon”. That promise has now been kept. The same team is behind Clef, the decision models we covered in October: Cloudflare Clef: Open-Source Decision Models for AI Agents.
A new architecture after the preview
Cloudflare says it fundamentally redesigned the architecture after the preview. Ordinary text and number fields are no longer produced in a single forward pass but through “parallel slot autoregressive generation”: each field uses the tokens it has already written to choose the next one and stops at its own end marker. All fields advance together in the serving batch. That costs a little latency, but according to Cloudflare it clearly improves accuracy on both text and images. The typed API with confidence scores and character spans remains unchanged.
PDF support is new. Each request can include up to four images or PDF pages in total. Invoice line items can come back as rows that keep description, quantity and price together, even when the supplier, line items and total appear on different pages.
The numbers: self-reported
Cloudflare publishes internal evaluations comparing the new models with its own preview, with Liquid AI’s D1.2 (paired with Apple Vision for optical character recognition) and with Scaledown’s Extract model. Some of the results:
- Fresh text fields (96 requests): Clairic got 335 of 336 fields right, Clairic-flash 328, the preview 295, D1.2 with OCR 304, Scaledown 259.
- Mixed-input fields (36 requests): Clairic 221 of 226, Clairic-flash 213, preview 207, D1.2 116, Scaledown 156.
- Tracked customer assertions (139 request shapes): Clairic and Clairic-flash 1,263 of 1,263 each, preview 1,033, D1.2 973, Scaledown 653.
- Synthetic receipt fields in text and images: both new models 576 of 576, preview 369.
On image reading, Cloudflare says errors on its internal benchmarks fell from 207 to zero. Latency on a self-hosted server stays below one second according to its own measurements: Clairic reaches 358 ms at the median and 736 ms at the 95th percentile on fresh text fields, Clairic-flash 155 and 328 ms. For image fields the figures are 453 and 976 ms and 305 and 605 ms respectively.
The important caveat comes from Cloudflare itself: there are no public extraction benchmarks yet, the results are self-reported and should be taken as directional guidance only. Perfect scores on internal test sets say little about how a model handles someone else’s documents. Independent testing is still outstanding.
Open weights built on Qwen
Both models are available on Hugging Face under the Apache 2.0 license, which also permits commercial use. According to the model cards, Clairic is based on Qwen3.8-27B and Clairic-flash on Qwen3.5-9B, both models from Alibaba. So Cloudflare did not train a base model from scratch but specialized open models for extraction. That does not diminish the practical value, but it puts the “superintelligence team” label into some perspective.
For local use, the model card lists Python 3.11 or later, PyTorch and Transformers; schema constraints are enforced by the xgrammar library through a logits processor. The bundled code is loaded through its own helper function, which Hugging Face flags as “custom code”. Anyone integrating the repository is therefore executing code from the model repository and should review it first.
Pricing: preview users will soon pay more
On Workers AI, the models are available as @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash. Only input tokens are billed. There is one detail developers should know: during the preview, the clairic endpoint ran the 9B model at 0.10 dollars per million input tokens. That model is now moving to clairic-flash, while clairic will serve the 27B model. Cloudflare is switching existing integrations over automatically.
Until December 1, 2026, both models cost 0.10 dollars per million input tokens. After that, the price for Clairic rises to 0.24 dollars, while Clairic-flash stays at 0.10 dollars. Anyone using the existing endpoint who changes nothing will therefore pay 2.4 times as much from December, in exchange for the larger model. For cost-sensitive workloads, Cloudflare itself recommends moving to Clairic-flash. According to the developer documentation, the API schemas remain unchanged.
Assessment
Clairic stands for a trend that makes fewer headlines than new frontier models but can save a lot of money in daily use: small specialist models that handle one task quickly and cheaply instead of sending every invoice field through a large language model. Meta is aiming at similar routine agent work with Muse Ember. The open weights under Apache 2.0 are a genuine plus, because they allow the models to run on your own hardware and with sensitive documents. The benchmarks, on the other hand, should be read with caution, since they come exclusively from Cloudflare. And anyone using the preview endpoint in production should decide before December 1 whether the larger model is worth the higher price.



