Cloudflare has begun developing its own artificial intelligence models, introducing Clef, a decision model, and Clairic, an extraction model. Both belong to the so-called System One category, designed for fast tasks with low latency. The company may eventually expand beyond that category.
Michelle Chen, who leads product for Cloudflare’s AI Platform, discussed the new direction at the AI Engineer Code Summit. In an interview with Latent Space, she said the company is also experimenting with reasoning models and other large language models. It is unlikely to focus on pre-training models from scratch in the manner of leading AI labs, however: most of the team works on post-training, with an emphasis on making existing resources more efficient and effective.
From infrastructure to model development
Chen described Cloudflare as a “neolab,” although the company only partly fits the term’s usual meaning. Stephanie Palazzolo of The Information coined it the previous November to describe a new generation of research-focused AI startups. According to Deedy Das, a partner at Menlo Ventures, such companies typically employ elite researchers, use substantial computing resources and raise significant funding, but have yet to generate substantial revenue. Examples include Thinking Machines Lab, Poolside and TypeSafe AI.
Cloudflare, by contrast, is a publicly traded company with significant revenue. Chief Executive Officer Matthew Prince said the company generated $782 million in third-quarter revenue, up 39% year over year, and passed $3 billion in annualized revenue.
Chen argued that model development is nevertheless a natural continuation of the company’s evolution. After infrastructure, Cloudflare built computing services – including its Workers platform and support for agents – and storage. Through Workers AI, it also serves open-weight models, including models from Kimi, GLM and DeepSeek.
Jev’s influence on Clef
The immediate inspiration for Cloudflare’s move into model development was TypeSafe AI’s Jev model, which arrived in September. Chen said the System One approach fits Cloudflare’s business because not every request needs to go to a large language model: smaller, faster models can handle some tasks, with response times measured in milliseconds rather than seconds.
Clef is a decision model built on the pattern established by Jev. Chen said Cloudflare built it in two weeks, whereas TypeSafe AI spent two years creating Jev. She credited the difference to TypeSafe AI’s pioneering work: Cloudflare already knew what approach to follow, including the structure of the API.
Clairic extracts values from unstructured data
Introduced last month, Clairic is Cloudflare’s own extension of the System One approach. As an extraction model, it can extract, calculate, estimate and forecast values from unstructured input.
The team looked for a task where large language models were excessive and did not always return the right answer. Chen said autoregressive language models can hallucinate or produce malformed JSON when processing unstructured data. She gave the example of an order number on an invoice: the information may be present in the document, yet the model can still return the wrong value.
Clairic is intended to extract a value of a specified type – a number, for example – from unstructured data in a single forward pass.
Reducing dependence on outside models
Chen said developing models in-house lets Cloudflare create capabilities it is missing and reduce risks in its AI operations. Jev was closed source, so Cloudflare could not run it itself; other models may present uncertainty through regulatory risks or expensive licences. The company is therefore exploring what additional models it could create from existing foundations, beyond its initial System One offerings.



