An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team
- ID
- 48042
- Status
- summarized
- Published
- 19 Nov 2026, 6:00 AM
- Fetched
- 19 Nov 2026, 9:06 PM
- Provider
- Cloudflare Blog
- Category
- infrastructure
- Original URL
- https://blog.cloudflare.com/open-clairic/
- Source URL
- https://blog.cloudflare.com/rss/
Summary
- Score
- 6.0
- Created
- 19 Nov 2026, 9:06 PM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_founders
What happened
Cloudflare released an updated Clairic plus a new Clairic-flash, a 9-billion-parameter lighter/faster variant, both open-weight on Hugging Face and hosted on Workers AI (announced 18 Nov 2026). The architecture changed since the preview: per-field extraction now uses parallel slot autoregressive generation instead of a single forward pass, each field stopping at its own end marker, and special heads handle estimates, forecasts, durations, calculated dates and arithmetic. New since preview is extraction from images and PDFs, with better table/list/date handling; the typed API keeps confidence scores, character spans and probabilities on every request. The post is truncated before the eval numbers, so no benchmark figures are given here.
Why it matters
If you currently pay a closed API per page to pull fields out of receipts, invoices or support tickets, this is a 9B open-weight model you can download and self-host or call through Workers AI, and because it still returns character spans and per-field confidence, you can route low-confidence fields to human review instead of writing regex post-validation. The decision point is cost and latency on your own document set, not the announcement — the excerpt cuts off before the evals, so benchmark it on your receipts/tickets before migrating.
Discussion angle
For a receipt or support-ticket extraction pipeline, does a task-specific 9B open-weight model with per-field confidence scores and character spans actually beat prompting a general frontier model — and what does self-hosting versus Workers AI do to your per-document cost?