Software Engineering

Cloudflare releases open-weight Clairic and Clairic-flash extraction models

Key Takeaway

Cloudflare has released open-weight versions of its redesigned Clairic extraction models, featuring a new parallel slot autoregressive architecture for high-speed data extraction from text, images, and PDFs.

Summary

Cloudflare's Workers AI superintelligence team has announced the release of open-weight versions of two specialized AI models: the redesigned Clairic and the compact Clairic-flash. These models are specifically engineered for high-accuracy data extraction from unstructured sources, including complex documents, images, and multi-page PDFs.

The updated Clairic architecture moves away from single forward passes in favor of parallel slot autoregressive generation, allowing multiple extraction fields to be processed simultaneously within a serving batch. While the full weights are available on Hugging Face for local deployment, the models are also hosted on Cloudflare Workers AI, providing a managed endpoint for developers requiring low-latency document processing.

Architectural changes in Clairic and Clairic-flash

  • Parallel Slot Autoregressive Generation: Unlike standard single-pass models, the new Clairic uses a method where each extraction field uses its own previously written tokens to select the next token, advancing together in a single serving batch to maintain low latency.
  • Specialized Computation Heads: The models include dedicated trained heads designed to handle specific mathematical and temporal tasks, such as arithmetic, duration calculations, forecasts, and estimated dates.
  • Multimodal Document Support: The architecture supports extracting structured data from up to four combined images or PDF pages, maintaining relational integrity for line items like descriptions, quantities, and prices.
  • Model Scaling and Parameters: The release includes the larger Clairic model and the Clairic-flash variant, which is a compact 9-billion-parameter model optimized for speed and efficiency.
  • API Metadata Retention: Despite the architectural redesign, the models continue to provide typed API outputs including confidence scores, character spans, and token probabilities.

Why it matters

By releasing open weights on Hugging Face, Cloudflare enables developers to run specialized, high-speed extraction models in private environments without relying on proprietary APIs. This provides a competitive open-source alternative for automated document processing, receipt scanning, and structured data ingestion in AI-driven workflows.

What to do

  • Download the open weights for Clairic or Clairic-flash from Hugging Face for local deployment.
  • Use the @cf/cloudflare/clairic-flash endpoint for the 9B model or the @cf/cloudflare/clairic endpoint for the 27B model on Workers AI.

Original reference: blog.cloudflare.com →