Cloudflare announced on November 18 that its document extraction models, Clairic and the newly introduced Clairic-flash, are now open-weight. The models are available on Hugging Face under the Apache 2.0 license and hosted on Workers AI, marking the graduation of a preview that launched in October. This move signals Cloudflareβs transition from a pure infrastructure provider to a frontier AI research lab, releasing weights that promise high-accuracy extraction for complex documents.
Architectural Overhaul and Capability Expansion
Since the October preview, Cloudflare has rebuilt the model to handle multi-page PDFs and complex layouts, moving beyond simple image reading. The new Clairic 27B utilizes parallel slot autoregressive generation, where each field uses its own tokens to predict the next, ensuring that line items, descriptions, and prices remain structurally linked. This architectural shift trades minor latency for significant accuracy gains, particularly in scenarios where extraction usually fails silently. The models also feature specialized heads for computation, allowing them to perform explicit arithmetic and generate forecasts directly from source data. Internal benchmarks claim image reading accuracy improved from 207 errors to zero, though these figures are self-reported. The system maintains strict grounding by returning confidence scores and character spans, allowing developers to verify that extracted values actually exist in the source document.
Two Models, Distinct Performance Profiles
Cloudflare has released two variants: Clairic 27B, based on Qwen3.8-27B, and Clairic-flash 9B, based on Qwen3.5-9B. The 27B model is optimized for maximum accuracy on hard documents, while the 9B flash model targets high-volume, low-latency tasks. In latency tests, Clairic-flash answers fresh text fields in 155ms at p50, compared to 358ms for the larger model. Both models operate well under one second, making them viable for synchronous request handlers rather than batch jobs.
Pricing Strategy and December 1 Deadline
Developers have a limited window to utilize the models at a promotional rate. Until December 1, 2026, both models cost $0.10 per million input tokens with no charge for output. After this date, the price for Clairic 27B rises to $0.24 per million input tokens, while Clairic-flash remains at $0.10. Users must note that the preview endpoint @cf/cloudflare/clairic now points to the 27B model; those requiring the cheaper 9B model must explicitly switch to the @cf/cloudflare/clairic-flash endpoint.
Limitations and Independent Verification
While the open weights are permissive, Cloudflare has not released its training data or full recipe, meaning this is not a fully reproducible open model. Furthermore, there are currently no public extraction benchmarks; all performance claims rely on Cloudflareβs internal evaluations against competitors like Liquid AI and ScaleDown. Independent testing is required to verify if the 27B model can truly outperform frontier models on diverse, real-world document layouts.
Key Takeaways
- Cloudflare released open weights for Clairic 27B and Clairic-flash 9B under Apache 2.0 on November 18.
- Pricing for the 27B model doubles from $0.10 to $0.24 per million input tokens on December 1, 2026.
- The 27B model uses parallel slot autoregressive generation for improved accuracy on multi-page PDFs.
- No public benchmarks exist yet; performance claims are self-reported by Cloudflare.
- Existing integrations using the preview endpoint will automatically switch to the new 27B model.
The Bottom Line
Cloudflare is no longer just a host for other people's weights; it is competing in the model lab space. If you are paying frontier prices for document extraction, this is a serious, cheaper alternative, but you must pilot it on your own edge cases before trusting the self-reported metrics.