During its “Birthday Week” on Oct 8 2026, Cloudflare announced Clef, a pair of open‑weight models built for decision‑making rather than text generation. The company made a 27‑billion‑parameter multimodal model and a smaller 9‑billion‑parameter version called Clef‑Flash available through Workers AI and as downloadable weights on Hugging Face.
How the models work
Clef takes a state description and a schema of typed questions as input and returns a probability for each allowed option. It can handle text, JSON, images or video in a single forward pass, so no extra parsing of free‑form output is needed. The 27B model has a 64 k token context window, twice what comparable models like Typesafe’s Jev provide.
Performance and access
Cloudflare says the 9B Clef‑Flash model achieves a median latency of 38.8 ms in its own benchmarks, while the larger 27B model records 209.3 ms. Because the models run on Cloudflare’s edge GPUs, the company expects low network latency for hot‑path decisions. Developers can fine‑tune the models with their own data using a new service that starts with engineer support and will later become self‑serve.
Why it matters
For developers building AI agents, Clef offers a ready‑to‑use way to make fast, structured decisions without writing custom parsers. The low latency of the 9B model makes it suitable for real‑time tasks like routing support tickets or filtering bot traffic, while the open‑weight release lets teams adapt the model to their own data.
