AI Agents That Can Actually Decide: Cloudflare Releases Clef
Most AI automation projects hit the same wall. The language model is smart enough to reason, but too slow and too expensive to put in the fast path of every decision your system makes. Routing a support ticket takes 500 milliseconds when you need 50. Classifying an incoming order for urgency burns tokens on a full generation pass when all you wanted was a yes or a no.
Cloudflare just released something built for that gap. It’s called Clef. It belongs to a new category of AI called decision models, and it works very differently from the large language models you’ve heard about.
A decision model is not a text generator. It doesn’t write explanations or produce paragraphs. You give it a situation and a set of questions with defined possible answers, and it returns typed outputs with probability scores attached. Send in a customer support message, ask whether it’s urgent and which team should handle it, and Clef responds with structured data your code can act on directly. No prompt engineering, no output parsing, no token waste.
That precision is what makes it fast. Clef-flash, the lighter version of the model, returns answers with a median latency of 38 milliseconds. Jev, the leading competing decision model, runs at 524 milliseconds. On a workflow making multiple classification calls per transaction, that difference adds up immediately.
Cloudflare tested Clef internally on their threat intelligence team, using it to classify website domains. The model determined whether a domain was a fashion site, an ecommerce store, a phishing attempt, or some combination. It finished in 2.2 seconds. Their fastest general-purpose language model took 4.7 seconds and returned fewer results. A real 2x speed gain on a real production workflow.
For a business owner, the benchmark numbers matter less than what they make possible. Put a fast AI classifier in the live path of your application without slowing it down. Every inbound support request gets triaged automatically before a human sees it. Incoming forms get validated for completeness and urgency in real time. Intent gets scored at scale, without noticeable delay.
Clef is open-source under the Apache 2.0 license and runs on Cloudflare’s Workers AI platform today. You can call it through an API, download the weights and run it locally, or work with Cloudflare’s team to fine-tune it on your own data. Fine-tuning matters. A model trained on your years of labeled decisions, resolved tickets, or historical transactions will outperform a generic one every time. That is the real advantage here.
Decision models are not replacing general AI. They fill a specific role, handling the parts of your workflow where you already know the questions and just need a fast, reliable answer. Putting that speed at the decision layer changes what automated systems can actually do.
Want to explore how AI decision models could benefit your business? Let’s talk.