Clef: Open-source decision models, and new RL fine-tuning platform

Over the past few weeks, there has been substantial interest in decision models like Typesafe AI's Jev System One. Unlike traditional classifier models, Jev introduces a new concept of decision models that produce bounded, structured outputs quickly and consistently. These models can handle any set of inputs without retraining, unlike Large Language Models (LLMs) that are non-deterministic and open-ended enough to reason and generate text for agentic workloads.

Today, Cloudflare is releasing two Cloudflare-trained decision models, Clef and Clef-flash, hosted on Workers AI. Clef is currently the leader when evaluated against the Jev Decision Index, and the models are fully Jev-API compatible, allowing easy experimentation. These models are smarter, faster, and open-sourced on Hugging Face under an Apache 2.0 license for local running and experimentation.

Decision models classify inputs into outcomes with probabilities. For instance, using Clef, a customer support message can be classified as urgent and the appropriate team to handle it. This means that a human is no longer always needed to make agentic decisions, as agents can gather context, make decisions, and take actions or defer to a human when necessary.

At Cloudflare, Clef has been tested on the Threat Intelligence team to classify website domains, resulting in faster and more accurate classifications than the fastest general LLM, gpt-oss-120b.

Clef has unique properties, such as a vision encoder for processing images and a larger 64k context window compared to Jev's 32k. It also scores competitively against other decision models on various benchmarks. Clef's performance is enhanced by being hosted on Workers AI, leveraging Cloudflare's GPUs at the edge for low network latency.

There are two Clef models: the larger Clef for powerful precision and the smaller Clef-Flash for latency-critical decisions. Both models are API-compatible and enterprise-ready, with a guarantee that Cloudflare does not read, store, or train on user requests or responses. Clef models can be used directly or fine-tuned for specific workloads through Cloudflare's fine-tuning services.