Cloudflare is moving from serving other companies’ AI models to building some of its own. Its first two System One models, Clef and Clairic, are designed for quick decisions and structured data extraction, rather than asking a large language model to do every job.
Cloudflare Watch analysis
What happened
At the AI Engineer Code Summit, Cloudflare AI Platform lead Michelle Chen described Clef as a decision model inspired by TypeSafe AI’s Jev, and Clairic as an extraction model. Chen said her team built Clef in two weeks, while crediting TypeSafe with the work behind Jev.
Clairic is intended to pull, calculate or estimate typed values from unstructured input in one forward pass. Chen gave invoices as an example: extracting an order number can be a better fit for a focused model than asking a general-purpose language model, which may return the wrong value or malformed JSON. In an interview with Latent Space, she said Cloudflare is also experimenting with reasoning models and other large language models, with post-training expected to be the main focus rather than training models from scratch.
Why it matters
Cloudflare already serves open-weight models through Workers AI. Building its own adds another step to that stack, and gives the company a way to tailor models to the low-latency work its infrastructure is built to handle. The proposed division of labour is straightforward: use a smaller, faster system for tasks it can handle, and reserve larger models for work that needs them.
That is also a business choice. Chen said developing its own models could reduce Cloudflare’s exposure to changes in model availability, licensing and regulation. The trade-off is that building and maintaining useful models takes effort; owning one is not, by itself, proof that it is better or cheaper.
Our read
The interesting move is not the “neolab” label. It is Cloudflare trying to make model-building another layer of its infrastructure business, beginning with tasks where speed and predictable output matter. Clairic’s extraction focus is a concrete example, not another promise to make a model that does everything.
The next test is whether these systems deliver dependable results in real customer workflows, and whether Cloudflare can support them alongside the models it already hosts. Smaller models earn their place by doing a specific job well, not by being small on a slide.
What to watch
- Which tasks Cloudflare assigns to Clef and Clairic in live products.
- Whether Cloudflare releases performance or reliability results for Clairic’s extraction work.
- Whether its planned reasoning and language models become products, and how they are licensed.
Discussion spark: Would you trust a focused model to extract important fields from invoices, or would you rather keep that job with a general-purpose model that can handle more kinds of input?
Sources and evidence
- Cloudflare Joins the Jevolution and Labels Itself a Neolab (17 November 2026, 15:00 UTC)
Independent WittyWires tracker for public updates about Cloudflare. Not affiliated with or endorsed by Cloudflare; this is not an official account.