
Cloudflare just made its Clairic data extraction models open-weight, so anyone can download and run them. The company released an updated 27-billion-parameter Clairic and a new, lighter 9-billion-parameter Clairic-flash, both hosted on Workers AI with full weights on Hugging Face. It also introduced the group behind them as the Workers AI superintelligence team.
Clairic is built for one job: pulling clean, typed fields out of messy inputs such as receipts, support tickets and documents. This update adds PDF support, a redesigned architecture and new pricing.
What Is New in Clairic Open-Weight Models
When Cloudflare first previewed Clairic, it read text in a single forward pass. The company says it went back to the drawing board after customer feedback and rebuilt the model around extraction. Here are the headline changes:
- Full open weights for both Clairic and Clairic-flash on Hugging Face.
- PDF reading, with up to four images or PDF pages per request.
- Line items that come back as rows, keeping description, quantity and price together.
- Better handling of tables, lists and dates, plus tighter grounding in the source text.
- Special heads for estimates, forecasts, durations, calculated dates and basic arithmetic.
A New Way to Generate Each Field
For ordinary text and number fields, Clairic now uses what Cloudflare calls parallel slot autoregressive generation. Each field writes its own tokens, uses what it already wrote to pick the next one, and stops at its own end marker. All fields move forward together in one batch, which keeps responses fast while improving accuracy.
The API still returns confidence scores, character spans and probabilities with every answer. Cloudflare says responses arrive in under one second, faster than typical large language models on similar jobs.
Pricing for the Open-Weight Clairic Models
Cloudflare is moving the original 9B preview model to a new Clairic-flash endpoint, while the bigger 27B model takes over the main Clairic name. Both charge only for input tokens, and Cloudflare absorbs the cost of output tokens.
| Model | Size | Workers AI endpoint | Price per million input tokens |
|---|---|---|---|
| Clairic | 27B parameters | @cf/cloudflare/clairic | $0.10 until Dec. 1, then $0.24 |
| Clairic-flash | 9B parameters | @cf/cloudflare/clairic-flash | $0.10 |
If your app already calls the preview Clairic endpoint, you get the new 27B model automatically, since Cloudflare is rolling it out in place. Cost-sensitive workloads can switch to Clairic-flash before the promotional price ends on December 1, 2026.
How Accurate Is It?
There are no public benchmarks for extraction yet, so Cloudflare shared its own internal evaluations. It says image reading errors on one internal benchmark dropped from 207 to zero after the redesign. The company compared Clairic with other extraction tools, such as Liquid AI’s D1.2 and Scaledown’s Extract, but it openly calls these results self-reported and directional.
Why Cloudflare Is Building Its Own Models
Cloudflare now describes itself as a “neolab” that builds frontier AI models for customers and the open-source community. The same team is behind Clef, an earlier model release, and says more models are on the way. Its pitch is simple: use small, fast, purpose-built models to replace costly large language model calls for narrow tasks.
The move fits a wider trend of companies shipping open-weight models that developers can tune and host themselves. Liquid AI and others are also chasing the extraction niche, which suggests real demand for cheaper ways to turn documents into data.
What This Means for You
If you process invoices, receipts, forms or support tickets, Clairic could cut both cost and latency compared with prompting a general chatbot. Here is how to approach it:
- Test the 27B Clairic now while it costs the same as Clairic-flash.
- Download the open weights if you need to run extraction on your own servers for privacy.
- Use the confidence scores and character spans to flag low-certainty fields for human review.
- Run your own tests, since the published accuracy numbers come from Cloudflare.
For background on the first release, see our earlier look at the Clairic preview launch.
Sources
Primary source: Cloudflare’s open-weight Clairic announcement.
Conclusion
Cloudflare’s updated Clairic and new Clairic-flash bring PDF reading, a rebuilt architecture and full open weights to its extraction models. The 27B model is cheap to try until December 1, and the 9B flash model stays at $0.10 per million input tokens. For teams that turn messy documents into structured data, this is a low-risk model worth testing against their current setup.


