Cloudflare Joins the Jevolution and Labels Itself a Neolab
- Source
- latent.space
- Author
- Richard MacManus
- Date

Context
Cloudflare, one of the web's essential infrastructure companies for the past decade and a half, now calls itself a neolab because it has started building its own AI models, Richard MacManus reports for Latent Space (latent.space). The claim came from Michelle Chen, who runs product for Cloudflare's AI platform, in a talk at AI Engineer Code Summit and an interview afterward. The word was coined by Stephanie Palazzolo of The Information for a new generation of research-focused AI startups, and a list of neolabs kept by investor Deedy Das describes them as heavily funded startups with elite researchers, large compute budgets and little revenue. Cloudflare misses much of that description: it is a public company that just reported quarterly revenue of $782 million, up 39% from a year earlier. MacManus concludes that the label pushes the accepted definition.
What has changed is that Cloudflare now releases models of a specific new kind. Its first, Clef, works with the same API as Jev, the model TypeSafe AI released in September. TypeSafe calls these System One models: instead of generating text, they return a structured answer, such as a choice among listed options with probabilities, in milliseconds. Chen said Cloudflare built Clef in two weeks against Jev's two years, and credited TypeSafe for inventing the category and settling what the API should look like. The second model, Clairic, applies the idea to extraction. It reads unstructured input such as an invoice and returns typed values, like an order number, in a single pass through the model. Chen's argument is that a writing its answer one at a time can waver over a number printed plainly on the page, or return malformed JSON.
Chen gave two reasons for building models. Sending some requests to a small, fast model instead of a language model fits a company whose core business is low latency. Owning models also reduces dependence on others: Jev was closed, so Cloudflare could not run it, and she pointed to possible regulatory risk around Chinese open models and to labs charging expensive licenses. She said reasoning models and other language models may follow, built by post-training existing models, since most of her team does post-training, rather than by training from scratch as frontier labs do. On that account, Cloudflare is an infrastructure company that has begun making specialized models to run on its own platform, which is a narrower thing than the frontier research the word lab usually suggests.
- Michelle Chen, who runs product for Cloudflare's AI Platform, said Clef, its decision model, took two weeks to build as a copy of TypeSafe AI's Jev, which took TypeSafe two years. She credited TypeSafe with inventing the category Cloudflare copied.
- Clairic, Cloudflare's extraction model, pulls typed values such as an invoice's order number out of unstructured input in one forward pass. Chen's rationale is that autoregressive LLMs can misread such details and sometimes return malformed JSON.
- Chen said Cloudflare may also ship reasoning models and other LLMs, but by post-training existing open models rather than from scratch, since most of her team does post-training.
- Chen framed owning models as derisking: Jev was closed source so Cloudflare could not run it, Chinese open models may face regulatory risk, and other model labs may enforce expensive licenses.
- Latent Space notes Cloudflare stretches the neolab label: Deedy Das's criteria include little revenue traction, while Cloudflare just reported $782 million in quarterly revenue, up 39% year over year, and crossed $3 billion annualized.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
articleThe Jev Revolution: What Engineers Are Already Building with System One Models
articleAI Engineer Code Summit: OpenAI on New Ways of Working With Codex
articleAgents Prompt People: Using Agents for Long-Horizon Financial Work
articleCan Frontier Models See Past a Decoy? Elorian Tests Visual Reasoning
Checking sign-in…
Loading comments…


