Skip to main content
Back

Cloudflare Opens Clairic Weights so Developers Can Self-Host Document Extraction

Cloudflare open-weights a redesigned 27B Clairic and new 9B Clairic-flash under Apache 2.0, adding PDF support and a parallel slot generation architecture tuned for typed data extraction.

Model, Full Access
  • Cloudflare released Clairic 27B and Clairic-flash 9B as open weights under Apache 2.0.
  • Both are multimodal extractors handling text, images and PDFs with typed field schemas and per-field confidence.
  • New architecture uses parallel slot autoregressive generation, with each field stopping at its own end marker.
  • Workers AI pricing: $0.24/M input for 27B, $0.10/M for flash, input-only billing, promo until Dec 1.
  • Clairic-flash hit 155ms p50 latency on fresh text extraction in Cloudflare's internal tests.
  • The 9B previously served at @cf/cloudflare/clairic now lives at @cf/cloudflare/clairic-flash.

Cloudflare has moved Clairic out of preview, released its weights under Apache 2.0, and added a smaller model called Clairic-flash. Both models are available for self-hosting through Hugging Face and as managed endpoints on Workers AI.

The release also introduces a redesigned decoder for typed fields. Clairic produces schema-constrained values, confidence scores, missing-value probabilities, and source locations, giving developers a specialized alternative to extracting JSON through a general-purpose chat model.

Schemas in, auditable fields out

Cloudflare groups Clairic under its “System One” label for models built around narrow tasks. An application supplies unstructured input, such as text, an image, or a PDF, along with a schema that defines the required fields. The model returns typed values for that schema.

The 27-billion-parameter Clairic supports text, numbers, dates, lists, and nested objects. Each field can include a confidence score and a probability that the requested information is absent. Text fields can also include character spans pointing to the relevant source content, which helps with invoice review, document audits, and human approval workflows.

A 9B fast path and a 27B accuracy tier

Clairic model lineup
Model Base model Workers AI ID Intended use
Clairic Qwen3.8-27B @cf/cloudflare/clairic Higher accuracy on difficult documents
Clairic-flash Qwen3.5-9B @cf/cloudflare/clairic-flash Lower latency and cost at higher request volumes

Cloudflare reassigned the endpoint used during preview. The 9B preview model previously ran at @cf/cloudflare/clairic; that identifier now serves the 27B model, while the 9B release has moved to @cf/cloudflare/clairic-flash. Existing callers of the original endpoint receive the larger model automatically, so teams should recheck latency, throughput, output behavior, and post-promotional pricing. Applications that need the previous size should switch to the flash identifier.

Each field gets its own decoder

Clairic now generates ordinary text and number fields through parallel slot autoregressive decoding. An autoregressive decoder predicts one token at a time using the tokens already generated. In Clairic, every requested field owns a separate decoding slot, and all slots advance together in the serving batch until each reaches its end marker.