Open & local AI

Cloudflare releases open-weight Clairic models for fast document extraction

Cloudflare launches updated Clairic and new Clairic-flash models with open weights, offering faster and more accurate data extraction from PDFs and images on Workers AI.

Documents turning into digital data streams entering a glass cube
Illustration generated for this article

Cloudflare has released updated versions of its Clairic extraction models, including a new lighter variant called Clairic-flash. The announcement, made on November 18, 2026, introduces open-weight models designed to improve data extraction from complex documents like PDFs and images. These models are now available on the Workers AI platform and Hugging Face, aiming to provide developers with faster and more cost-effective tools for structured data retrieval.

What happened

The Workers AI superintelligence team at Cloudflare introduced two specific models to address common pain points in document processing: accuracy, latency, and cost. The first is an updated version of Clairic, which has been redesigned since its preview release to handle unstructured data more effectively. The second is Clairic-flash, a compact 9-billion-parameter model optimized for speed while maintaining strong extraction performance. Both models are now open-weight, allowing the community to inspect, fine-tune, and deploy them independently of Cloudflare’s infrastructure if desired.

This release follows significant feedback from users who tested the initial preview of Clairic. Developers reported needs ranging from extracting values from support tickets to reading line items from receipt images. In response, Cloudflare re-architected the model to focus exclusively on extraction tasks rather than general-purpose language generation. The team also highlighted that these models can now process PDFs directly, a capability that was absent in the preview version. This allows users to request supplier names, line items, and totals from multi-page documents in a single API call.

To encourage adoption, Cloudflare is offering a promotional pricing structure until December 1, 2026. During this period, both the larger Clairic model and the smaller Clairic-flash model will cost $0.10 per million input tokens. After this date, the price for the larger 27-billion-parameter Clairic model will rise to $0.24 per million input tokens, while Clairic-flash will remain at $0.10. The company emphasizes that it absorbs the cost of output token generation, adhering to its System One pricing principles.

How it works

The core technical improvement in the new Clairic architecture is the use of parallel slot autoregressive generation. Unlike previous approaches that might attempt to extract all fields in a single forward pass or sequentially, this method allows each field to generate tokens independently but simultaneously. Each field uses the tokens it has already written to determine the next token, stopping only when it reaches its own end marker. Because all fields advance together within the serving batch, the system maintains low latency while significantly improving accuracy for both text and image inputs.

For tasks involving computation or estimation, such as forecasting durations or calculating dates, the models employ special training heads. These heads are designed to handle explicit arithmetic and logical estimations directly, reducing the need for post-processing scripts. Despite these architectural changes, the API interface remains consistent with the preview version. Developers still receive confidence scores, character spans, and probabilities for every extracted field, ensuring that applications can assess the reliability of the data before using it.

Clairic-flash serves as a streamlined alternative for use cases where speed is critical. As a 9-billion-parameter model, it delivers responses significantly faster than the larger variant. Benchmarks show that Clairic-flash can process fresh text fields with a median latency of 155 milliseconds, compared to 358 milliseconds for the full Clairic model. This makes it suitable for high-throughput environments where sub-second response times are mandatory, such as real-time expense auditing or traffic forecasting.

Key details

  • Open-weight availability: Both Clairic (27B parameters) and Clairic-flash (9B parameters) are available as open-weight models on Hugging Face.
  • PDF support: The updated models can process up to four images or PDF pages per request, extracting structured data like line items across multiple pages.
  • Pricing promotion: Until December 1, 2026, both models cost $0.10 per million input tokens on Workers AI; thereafter, Clairic rises to $0.24 per million input tokens.
  • Performance gains: Internal evaluations show Clairic achieving perfect scores on synthetic receipt fields (576/576), eliminating errors seen in the preview version.
  • Latency benchmarks: Clairic-flash achieves p50 latencies under 200ms for many tasks, such as loan pricing (188ms) and traffic forecasting (243ms).
  • API consistency: The API schema remains unchanged from the preview, preserving features like confidence scores and character span grounding.

Why it matters

For software engineers building document-intensive applications, the shift toward specialized extraction models represents a move away from expensive, general-purpose large language models. Traditional LLMs often struggle with structured output, requiring complex prompt engineering and post-processing to ensure data integrity. Clairic’s design, which focuses solely on extraction, reduces this overhead. By providing confidence scores and character spans, it allows developers to build robust validation layers without sacrificing speed. This is particularly valuable for industries like finance and logistics, where accurate data entry from invoices and receipts is critical.

The open-weight nature of these models also lowers the barrier to entry for self-hosted solutions. Teams with strict data privacy requirements can download the models from Hugging Face and run them on their own infrastructure. This flexibility complements cloud-based offerings like Workers AI, giving organizations the choice between managed convenience and local control. Furthermore, the availability of a lightweight variant like Clairic-flash enables edge computing scenarios where resource constraints previously made AI-driven extraction impractical.

What you can do

  • Test the new Clairic and Clairic-flash endpoints on Workers AI using the existing API schema to evaluate accuracy on your specific document types.
  • Download the open-weight models from Hugging Face if you need to self-host for compliance or customization purposes.
  • Migrate cost-sensitive workloads to Clairic-flash to take advantage of its lower latency and permanent $0.10 per million input token pricing.
  • Implement validation logic using the provided confidence scores and character spans to filter out low-certainty extractions in production.
  • Take advantage of the promotional pricing before December 1, 2026, to benchmark the higher-quality 27B Clairic model at the reduced rate.
  • Explore the support for PDF inputs by sending multi-page documents to test the model’s ability to link line items across different pages.
  • #cloudflare
  • #ai
  • #document-extraction
  • #open-source
  • #workers-ai

Tools from the Bytechap store

Keep reading

All stories