Skip to content
INTERNET INFRASTRUCTURE & NETWORKS

Cloudflare Positions Itself as a "Neolab" With Major Open-Weight Releases of Frontier AI Extraction Models

Cloudflare is officially stepping into the role of a "neolab," signaling a major strategic shift toward building its own frontier artificial intelligence models directly for customers and the broader open-source community. As part of this evolution, the company is reintroducing its core AI research group as the Workers AI superintelligence team. This specialized unit, operating behind notable projects like Clef and Clairic, is moving rapidly to push the boundaries of frontier AI and deliver practical, high-performance models tailored to real-world infrastructure and data extraction demands.

Today, the team is making significant waves across the tech industry by announcing two open-weight model releases: an upgraded flagship Clairic model and a new, faster variant called Clairic-flash. Both specialized extraction models are hosted directly on Cloudflare’s Workers AI platform, while their full open weights have been made publicly available on Hugging Face. This move underscores Cloudflare’s broader mission to help organizations augment their expensive, general-purpose large language model (LLM) workflows with smaller, faster, and dramatically cheaper models purpose-built for specific, high-frequency enterprise tasks.

A Redesigned Architecture for Clairic

The journey to today’s release began when Cloudflare first previewed its Clairic extraction model. Following that initial rollout, the engineering team received a tremendous volume of real-world use cases and constructive feedback from customers detailing how they wanted to leverage an automated extraction tool. These requirements spanned a wide spectrum of applications, ranging from pulling structured line items out of image-based receipts to parsing critical data fields from complex support tickets and customer service logs. Through this feedback, Cloudflare recognized that extraction had to remain the unambiguous focus of the model, meaning it needed to excel above all at reading and structuring messy, unstructured data.

Rather than sticking with its original approach, Cloudflare decided to go back to the drawing board, completely redesigning the underlying architecture to differentiate this new production version from the initial preview release. For standard text and numeric extraction fields, the team transitioned away from a single forward pass, adopting a technique known as parallel slot autoregressive generation instead. Under this new method, each target field uses the tokens it has already written to intelligently choose its next token, continuing until it encounters its own designated end marker. All of the fields advance simultaneously within the serving batch, ensuring that the system still delivers lightning-fast latency per request.

This architectural pivot yielded a substantial boost in extraction accuracy for both raw text and complex image inputs, effectively offsetting the minor trade-off in processing latency. Furthermore, to address specialized computational requirements, the team trained dedicated model heads capable of handling estimates, financial forecasts, durations, calculated dates, and explicit arithmetic operations.

Despite these deep architectural evolutions, Cloudflare consciously preserved the vital components of the typed API interface that made the initial tool so useful for developers. Features such as confidence scores, precise character spans, and probability metrics remain fully intact with every incoming request. The models continue to operate with remarkable speed and cost-efficiency. Internal evaluations demonstrate that these newer versions of Clairic have vastly improved extraction performance across plain text, high-resolution images, multi-page PDFs, and other complex document formats.

While independent, standardized public benchmarks for dedicated extraction models remain scarce, Cloudflare shared internal evaluation data comparing its models against other emerging extraction solutions on the market, such as Liquid AI’s D1.2 model paired with Apple Vision for optical character recognition, and Scaledown’s Extract tool. Although Cloudflare notes that these self-reported evaluations should be viewed as directional guidance, they provide an illustrative picture of how Clairic performs across generalized extraction tasks. More importantly, the existence of competing extraction offerings from other companies validates Cloudflare’s original hypothesis: that data extraction has long been a major pain point for LLM users due to prohibitive costs, high latency, and inconsistent accuracy.

Measuring Performance and Processing Speed

In real-world applications, performance is defined just as much by speed as it is by accuracy. To verify the responsiveness of the redesigned models, Cloudflare ran comprehensive latency tests measuring complete HTTP responses from self-hosted servers, encompassing every requested field in a single payload. Even with the more sophisticated parallel slot autoregressive architecture in place, users can consistently expect complete responses in under one second—substantially faster than typical responses generated by standard general-purpose LLMs handling comparable workloads.

Testing across diverse enterprise workloads highlighted this speed advantage. For fresh text field evaluations spanning nearly a hundred requests, the standard Clairic model achieved a median (p50) latency of 358 milliseconds and a 95th percentile (p95) latency of 736 milliseconds, while the lighter Clairic-flash variant dropped those numbers down to a blistering 155 and 328 milliseconds, respectively. When processing complex commercial lease documents, traffic forecasting data, loan pricing sheets, and financial expense audits, both models maintained sub-second response times across the board, proving

An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team

their viability for latency-critical production pipelines.

Enhanced Capabilities for Images and PDFs

When Cloudflare first introduced the Clairic preview, its primary demonstration involved reading a photo of a single receipt. With today’s official release, the expanded Clairic family of models gains native support for parsing multi-page PDF documents. Users can now submit a request asking for a supplier’s name, individual line items, and final totals all at once, even when that information is spread across entirely different pages within a document. Applications can include up to four images or PDF pages in a single request, and complex line items can be returned cleanly as structured rows that keep descriptions, quantities, and prices securely aligned.

Driven by the structural improvements in the new architecture, image reading accuracy has seen a dramatic leap forward. On Cloudflare’s internal benchmarks covering mixed text and image extraction, error rates plummeted from over two hundred errors down to zero, highlighting the reliability of the updated parsing engine when handling messy real-world visual data.

Promotional Pricing Structure and Model Tiers

To accommodate different enterprise needs and budgetary constraints, Cloudflare is splitting its offerings into two distinct tiers. The original 9B parameter variant, which was favored during the preview phase for its rapid training and deployment profile, is moving to the @cf/cloudflare/clairic-flash endpoint. Meanwhile, the more powerful 27B parameter model will inhabit the primary @cf/cloudflare/clairic endpoint.

Under the permanent pricing structure, the larger 27B Clairic model will be billed at $0.24 per million input tokens, while the 9B Clairic-flash model will cost $0.10 per million input tokens. Sticking to the core economic principles established by Cloudflare’s System One initiative, the company has decided to absorb the cost of output token generation entirely. Organizations are charged exclusively on input tokens, significantly simplifying cost forecasting for developers.

To ensure a seamless transition for existing users, Cloudflare is introducing a temporary promotional discount. Running until December 1, both the standard Clairic and Clairic-flash endpoints will be priced uniformly at $0.10 per million input tokens. This promotional window allows developers currently hooked up to the original preview endpoint to automatically benefit from the vastly improved extraction capabilities of the 27B model without experiencing any price increase. Developers managing highly cost-sensitive workloads can deliberately route their traffic to the lighter 9B Clairic-flash variant at any time. When the promotion concludes on December 1, pricing for the flagship Clairic endpoint will adjust to its standard rate of $0.24 per million input tokens.

Open-Weight Availability and Continued Community Engagement

Both Clairic models are fully available as open-weight releases on Hugging Face, reaffirming Cloudflare’s commitment to supporting the broader open-source research and development community. The company witnessed an enthusiastic wave of community experimentation and fine-tuned models following its earlier release of the Clef model, and leadership is eager to foster that same collaborative momentum with Clairic.

The overarching mission behind Cloudflare’s System One model family remains focused on empowering organizations to supplement expensive general-purpose LLM usage with smaller, faster, and more economical tools designed explicitly for narrow, high-value tasks. By publishing the full open weights, Cloudflare hopes to encourage engineers and researchers worldwide to experiment with, build upon, and push the boundaries of these novel extraction architectures.

Both Clairic and Clairic-flash are live on Workers AI under their respective endpoints. Complete API schemas and technical documentation remain unchanged from the initial preview release, ensuring that current integrations continue to function smoothly. As the Workers AI superintelligence team continues to expand its footprint and accelerate its research initiatives, Cloudflare is actively recruiting engineering talent to join the group early and help shape the future trajectory of the company’s frontier AI ecosystem.