Clairic Extraction Models: Cloudflare Unlocks 2x AI Power

Clairic extraction models - Clairic Extraction Models: Cloudflare Unlocks 2x AI Power

Cloudflare Unveils Open-Weight Clairic Extraction Models, Boosting AI Capabilities

Cloudflare has officially entered the frontier AI research arena with the release of open-weight versions of its advanced extraction models, Clairic and Clairic-flash. Developed by the newly formed Workers AI Superintelligence team, these models are poised to significantly enhance how businesses and developers extract critical information from complex documents and images. This move democratizes access to powerful AI tools, fostering innovation within the open-source community and for Cloudflare’s extensive customer base.

The updated Clairic model showcases substantial improvements since its preview, particularly in handling complex documents that include images and PDFs. Its ability to reliably extract tables, lists, and dates has been refined, ensuring better grounding of answers within the source material. Complementing this is Clairic-flash, a more compact and faster 9-billion-parameter model that still delivers robust extraction and prediction performance, making it ideal for cost-sensitive and latency-critical applications.

Revolutionary Architecture for Enhanced Data Extraction

The Workers AI Superintelligence team has redesigned Clairic’s architecture from the ground up, moving beyond a single forward pass for extraction. The new approach utilizes parallel slot autoregressive generation, where each extraction field independently processes tokens and determines its own end point. This parallel advancement across all fields within a serving batch dramatically improves accuracy for both text and image data, with only a minor trade-off in latency.

Specialized heads have been trained to handle a variety of complex tasks, including estimations, forecasts, durations, calculated dates, and arithmetic operations. Despite these architectural enhancements, Clairic retains its user-friendly typed API interface, providing essential confidence scores, character spans, and probabilities with every output. This focus on practical utility ensures that the Clairic extraction models remain both incredibly fast and cost-effective for a wide range of applications.

Performance Benchmarks and Competitive Edge

Internal evaluations demonstrate Clairic’s superior performance compared to other market-leading extraction models, including Liquid AI’s D1.2 and Scaledown’s Extract. The updated Clairic model achieved perfect scores on internal benchmarks for text and image extraction, a significant leap from its preview version. Clairic-flash also shows remarkable efficiency, often matching the accuracy of the larger Clairic model while operating at substantially lower latency.

Latency tests reveal that both models deliver complete HTTP responses in under one second, a critical advantage over typical LLM responses for similar tasks. For instance, the Clairic-flash model achieves p50/p95 latencies as low as 155/328 ms for fresh text fields, while the larger Clairic model handles commercial leases with p50/p95 latencies of 691/803 ms. These figures highlight the engineering prowess behind the Clairic extraction models.

Accessibility, Pricing, and Future Outlook

Both Clairic and Clairic-flash are now available as open-weight models on Hugging Face, encouraging broader experimentation and development. Cloudflare is offering promotional pricing until December 1, 2026, making both models available at $0.10 per million input tokens. This allows users to leverage the enhanced capabilities of the larger Clairic model at the price point previously reserved for Clairic-flash.

Post-promotion, the 27B Clairic model will be priced at $0.24 per million input tokens, while the 9B Clairic-flash will remain at $0.10 per million input tokens. Cloudflare’s commitment to charging only for input tokens, absorbing output token generation costs, further enhances the value proposition of these Clairic extraction models. The Workers AI Superintelligence team is actively expanding its research, promising further advancements in frontier AI development.

Engineering Trade-offs and Limitations

While the new architecture significantly boosts accuracy, the parallel slot autoregressive generation introduces a slight latency increase compared to a hypothetical single-pass model, though it remains exceptionally fast. The models are optimized for extraction tasks, meaning their performance on generative or conversational AI tasks may not be as competitive as larger, more general-purpose LLMs. Cloudflare also notes that while they have run internal evals against other emerging models, public extraction benchmarks are still scarce, making direct, standardized comparisons challenging.

The current iteration supports up to four images or PDF pages per request, which may limit very large document processing in a single call. Furthermore, while Clairic excels at structured data extraction from unstructured sources, highly nuanced or subjective content interpretation might still require human oversight. The team acknowledges that these Clairic extraction models are part of an ongoing research effort, with continuous improvements planned based on user feedback and evolving AI research frontiers.

Field Readiness and Verdict

Cloudflare’s Clairic and Clairic-flash represent a significant leap forward in accessible, high-performance AI for data extraction. Their open-weight nature, coupled with competitive pricing and impressive accuracy, makes them a compelling choice for developers and businesses seeking to automate document analysis and information retrieval. The models are readily available on Workers AI and Hugging Face, ensuring ease of integration.

The Workers AI Superintelligence team’s dedication to pushing the AI frontier is evident in these releases. The Clairic extraction models are not just tools but platforms for innovation, enabling new applications in areas like automated customer support, financial document processing, and intelligent data ingestion. For organizations looking to harness the power of AI for extraction tasks, Cloudflare’s latest offerings are a must-evaluate, promising enhanced efficiency and reduced operational costs.

Source: Cloudflare Engineering Blog