Cloudflare announced a new step in its AI development path with the release of two versions of its data extraction models under an open-weight system: the updated Clairic model and the new, faster Clairic-flash model. This step comes as part of the company's strategy to empower developers to use specialized and lower-cost models compared to general Large Language Models (LLMs).
Specialized models and improved architecture
The Clairic model is available in a version with 27 billion parameters, and it has undergone a comprehensive redesign of its architecture. Instead of relying on a single forward pass, the model now uses parallel slot autoregressive generation technology to extract text and numeric fields. This mechanism allows each field to use tokens written previously to select the next token, significantly enhancing accuracy while maintaining high response speed.
As for the Clairic-flash model, it is a lighter and faster version comprising only 9 billion parameters, designed specifically to deliver strong performance in extraction and prediction tasks with lower resource consumption, making it ideal for applications requiring ultra-high speed.
What changes practically?
This update represents a shift in how unstructured data is handled, highlighting the following benefits:
- Support for complex documents: The Clairic family of models can now read PDF files and images with high accuracy, with the ability to extract data from multiple pages in a single request (such as extracting supplier name, line items, and totals from an invoice spread across different pages).
- Data extraction accuracy: Internal tests showed a massive improvement in image reading accuracy, with errors in some tests dropping from 207 errors to zero.
- Speed and efficiency: The models provide a response time of less than one second, vastly outperforming traditional large language models when executing extraction tasks.
- Handling calculations: Special heads in the model were trained to handle estimations, forecasts, durations, and explicit arithmetic operations.
Availability and pricing
Developers can access these models via Cloudflare's Workers AI platform, and the open weights are also available for download via the Hugging Face platform. To encourage experimentation, the company introduced a uniform promotional price of $0.10 per million input tokens for both models until December 1, 2026. After this date, the price of the Clairic (27B) model will revert to $0.24 per million input tokens, while the price of Clairic-flash (9B) will settle at $0.10.