Cloudflare is positioning itself as a research lab, developing frontier artificial intelligence models intended for both its customer base and the broader open-source ecosystem.

The company announced two open-weight model releases today. An updated iteration of Clairic incorporates improvements to extraction capabilities from complex documents, including images and PDFs, with enhanced accuracy for tables, lists, and dates, plus stronger grounding of extracted information in source material. Clairic-flash represents a streamlined, faster alternative built on a 9-billion-parameter architecture that maintains strong extraction and prediction performance while reducing computational overhead.

Both models are accessible through Workers AI, with full open weights published on Hugging Face. Cloudflare simultaneously rebranded its frontier AI research division as the Workers AI superintelligence team, the group responsible for developing Clef, Clairic, and additional models in development. The team is focused on advancing AI capabilities while creating practical models for Cloudflare users and the open-source community.

Architectural redesign and performance improvements

Following the preview release, Cloudflare gathered extensive customer feedback on extraction use cases spanning receipt image processing to data extraction from support tickets. This input prompted a complete architectural overhaul to position extraction as the model's primary strength when working with unstructured data.

The redesigned Clairic employs parallel slot autoregressive generation for text and number extraction fields rather than a single forward pass. Each field independently uses previously generated tokens to determine its next token and terminates at its own end marker. Fields progress together within the serving batch, maintaining rapid latency per request. This approach improved extraction accuracy for both text and images, with the accuracy gains justifying minimal latency trade-offs. For computation, estimation, and extraction scenarios, the model includes specialized heads capable of handling estimates, forecasts, durations, calculated dates, and explicit arithmetic operations.

The typed API interface retained its most valuable features: confidence scores, character spans, and probabilities accompany every response. The model maintains its speed and cost efficiency. Internal evaluations demonstrate substantial improvements in extraction performance across text, images, PDFs, and additional formats compared to earlier versions.

Cloudflare ran internal evaluations against emerging extraction models including Liquid AI's D1.2 (combined with Apple Vision for optical character recognition) and Scaledown's Extract. Clairic models demonstrated superior performance across generalized extraction scenarios, though Cloudflare notes these self-reported evaluations should be treated as directional guidance. The company highlighted that competing releases from other vendors validate extraction as a genuine pain point for LLM users due to cost, latency, and accuracy concerns.

Latency performance

Latency testing for both Clairic and Clairic-flash measured complete HTTP responses from a self-hosted server, encompassing all fields in requests. Despite architectural changes, responses arrive in under one second—substantially faster than typical LLM response times for comparable workloads.

Enhanced image and PDF handling

The preview version of Clairic demonstrated receipt image reading capabilities. The updated Clairic family now processes PDFs as well, enabling extraction of supplier information, line items, and totals in a single request even when distributed across multiple pages. Users can submit up to four images or PDF pages combined, with line items returned as structured rows preserving description, quantity, and price associations. Image reading accuracy improved from 207 errors to zero on internal benchmarks covering text and image content.

Pricing and promotional offer

clairic-image-extraction-comparison.png
Side-by-side extraction results for a synthetic coffee shop receipt. Clairic-preview returned null for the merchant name, while Clairic returned “Coffee House” and got all five fields correct.

The initial preview endpoint for Clairic at @cf/cloudflare/clairic hosted a 9-billion-parameter variant priced at $0.10 per million input tokens. Cloudflare is now relocating the 9B model to the @cf/cloudflare/clairic-flash endpoint, with the 27-billion-parameter model residing at @cf/cloudflare/clairic.

Standard pricing places the 27B Clairic model at $0.24 per million input tokens and the 9B Clairic-flash model at $0.10 per million input tokens. Although the model generates output tokens, Cloudflare applied System One pricing principles and charges only on input tokens, absorbing output token generation costs.

To facilitate a smooth transition, Cloudflare is offering promotional pricing through December 1, 2026: both Clairic and Clairic-flash will cost $0.10 per million input tokens. This allows users to evaluate the higher-quality Clairic variant at the same price point as the preview offering. Applications connected to the preview Clairic endpoint will automatically receive the improved extraction capabilities of the new model as it rolls out. Users operating cost-sensitive workloads can migrate to the more economical 9B Clairic-flash variant. On December 1, promotional pricing expires and Clairic pricing will increase to $0.24 per million input tokens.

Open-weight availability

Both Clairic variants are now open-weight on Hugging Face. Following the open-weight release of Clef, which generated numerous experimental and tuned model variants, Cloudflare continues supporting research teams and practitioners through its open-weight Clairic release.

The System One model initiative aims to help users supplement expensive LLM usage with smaller, faster, cheaper models optimized for specific tasks. Cloudflare expects the open-weight release will inspire additional experimentation and development building on its approach.

Accessing the models

Clairic and Clairic-flash are accessible on Workers AI via @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash respectively. Complete API schemas are available in developer documentation, with interfaces unchanged from the preview endpoint. Open-weight releases for both models are published on Hugging Face.

The Workers AI superintelligence team is in early stages of expansion and actively recruiting to advance frontier AI research at Cloudflare. The company is inviting early adopters to participate and influence the team's direction.