DEVOURED
Cloudflare's Clairic and Clairic-flash
Cloudflare has released open-weight versions of its extraction models, Clairic and Clairic-flash, designed for structured data extraction from text, images, and PDFs.
What: Cloudflare's new 'Workers AI superintelligence team' released the 27B-parameter Clairic and the 9B-parameter Clairic-flash on Hugging Face. The models use a parallel slot autoregressive architecture to improve accuracy in data extraction, specifically handling PDFs and images better than the preview release.
Why it matters: By releasing open-weight models, Cloudflare is attempting to create a 'System One' ecosystem where developers move specialized extraction tasks away from expensive, generalized LLMs toward smaller, cheaper, and faster purpose-built models.
Takeaway: Developers can now run these models on Cloudflare Workers AI or self-host them via the Hugging Face weights.
Deep dive
- Redesigned architecture: Shifted from a single forward pass to parallel slot autoregressive generation, allowing fields to be generated concurrently for higher accuracy.
- Capability expansion: Added support for PDF and image inputs in addition to raw text.
- Pricing strategy: Transitioning to a model where developers are charged only on input tokens, with Cloudflare absorbing the cost of output generation.
- Performance: Models report latency under one second for most tasks on Workers AI.
- Benchmark status: Internal benchmarks report zero errors on image/text extraction for the new models, though Cloudflare notes these are not yet externally verified.
Decoder
- Open-weight: Models where the trained weights are publicly available for download and use, though they may lack the full training dataset or exact training pipeline transparency of 'Open Source' models.
- Parallel slot autoregressive generation: A technique where multiple fields in a template are generated simultaneously in a single pass to reduce latency while maintaining contextual awareness.
- Workers AI: Cloudflare's serverless AI inference platform that allows developers to run inference on their global network.
Original article
Full article content is not available for inline reading.
Explore all 56 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi