Cloudflare just launched Clef, its first in-house AI models, and says they beat Jev on speed. Here's what you need to know. Clef is the first model family trained by Cloudflare's Workers AI team. There are two decision models, Clef and Clef-flash. Cloudflare says both top the benchmarks for quality and latency. It also launched a new RL fine-tuning platform alongside them. Decision models don't write text. They take a state and a schema of typed questions and return decisions, for routing, sorting and classification. Clef is a 27B multimodal model built on a Qwen backbone. It uses a non-autoregressive, prefill-only scoring approach. Jev, from TypeSafe AI, is the model that started this wave. Cloudflare says Clef leads the evals on the Hugging Face Jev Decision Index. It claims it pushes the pareto frontier on both quality and latency. A live leaderboard and demo are up on a Workers AI site. Clef is hosted on Workers AI and the weights are on Hugging Face. They are open source under Apache 2.0, so you can run them locally. One early user already ran clef-flash on an M5 Max. Cloudflare also says Replicate's Cog work is now part of how you fine-tune and redeploy models on Workers AI. Key numbers: - clef-flash is around 13x faster than Jev - Trained for 256k context, 64k hosted - Vision encoder, so it classifies images as well as text - Drop-in compatible with the Jev API - License: Apache 2.0 The race is on, and Cloudflare is the first big infrastructure player to ship open decision models that claim to beat Jev.