Cloudflare is moving from serving other companies’ models to building its own narrow AI systems. Its first two releases, Clef and Clairic, target decisions and structured extraction that the company says can run in milliseconds rather than sending every request to a large language model.

Clef follows the approach of TypeSafe AI’s Jev decision model. Cloudflare says it built its version in two weeks after the earlier system demonstrated the pattern. Clairic applies a related idea to unstructured inputs, extracting typed values such as an invoice number in one forward pass instead of generating an open-ended response that may contain malformed JSON or invented details.

The strategy fits Cloudflare’s latency-focused network and Workers platform. Smaller “System One” models can handle constrained steps while larger models remain available for tasks that need richer reasoning. Michelle Chen, who leads product for Cloudflare’s AI platform, said the team is also experimenting with reasoning models and other LLMs, though it does not plan to pretrain frontier models from scratch.

Cloudflare calls this expansion a move toward becoming a new kind of AI lab, but the label is debatable: it is a profitable infrastructure company rather than a research startup. The practical change is clearer than the branding. Cloudflare can now build specialized components when general models are too slow, expensive or unpredictable for a network-scale workflow.