Cloudflare, long known for speeding up the web, has moved into building its own AI models. At the AI Engineer Code Summit, Michelle Chen, who runs Cloudflare's AI Platform on the product side, presented two 'System One' models: Clef, a decision model modeled on TypeSafe AI's Jev, and Clairic, which Cloudflare describes as a new class of 'extract' model launched last month. The models are designed to handle work that does not need a large language model, with outputs measured in milliseconds rather than seconds. Chen said Clairic targets unstructured input, where LLMs can hallucinate or return malformed JSON.
Chen told Latent Space the company will not stop at System One models. She said reasoning models and other LLMs are under experimentation, though she does not expect Cloudflare to pre-train models from scratch. Instead, the focus will be post-training existing resources, since most of the team works on post-training. Cloudflare already serves open-weight models such as Kimi, GLM and DeepSeek through Workers AI. Chen said Clef took about two weeks to build, crediting TypeSafe AI, which spent roughly two years developing Jev.
Whether Cloudflare qualifies as a 'neolab' is debatable. The term, coined by Stephanie Palazzolo of The Information, describes research-focused AI startups that employ elite researchers, spend heavily on compute and lack significant revenue. Cloudflare is a public company that reported record third-quarter revenue of $782 million, up 39% year over year, and crossed $3 billion in annualized revenue. Chen framed the model push as a natural progression from infrastructure to compute, storage and now models. For Australian businesses and professionals, the direct impact is limited for now, since these models are aimed at developers building on Cloudflare's platform. Teams already using Workers AI could test a small, fast model for pulling structured data from messy text such as emails, invoices or forms, which may lower inference costs and latency compared with a frontier LLM. Any such trial should be checked against the Privacy Act 1988 and the Australian Privacy Principles, because customer data sent to a third-party model remains personal information. Otherwise, there is nothing to act on yet.
Key Points
Cloudflare has launched two System One models, Clef and Clairic, aimed at fast decisions and data extraction.
Clef is a copy of TypeSafe AI's Jev; Chen said it took about two weeks to build, versus about two years for Jev.
Clairic is a new 'extract' model class that pulls and estimates values from unstructured input.
Cloudflare plans reasoning models and other LLMs, focusing on post-training existing open models rather than pre-training from scratch.
The 'neolab' label is a stretch, since Cloudflare is a public company with $782 million in quarterly revenue.