Cloudflare launched the Clef and Clef flash, its first proprietary models focused on structured decisions for AI agents. Available on Workers AI and with weights published under the Apache 2.0 license, they were developed to classify requests, choose action routes and determine when an agent should escalate a decision to a person.
Announced on Thursday (1st), the technology targets a different stage from the one occupied by large language models. Instead of producing open text, Clef receives a defined set of options and returns probabilities for each permitted answer, generating a structured output that can be consumed directly by software.
In practice, an agent can use the model to identify urgency, forward a request to the correct team or decide when human intervention is needed. The proposal is to remove recurring classification tasks from the path of larger LLMs, reducing latency and computational cost.
Clef bets on speed for structured decisions
Cloudflare states that the Clef achieved the best score in seven of ten decision benchmarks evaluated by the company. In latency tests, the model recorded a median of 209.3 milliseconds, while Clef flash reached 38.8 milliseconds. Jev, used as a reference, recorded 524.1 milliseconds.
Clef also offers image classification and a context window of 64 thousand tokens. According to the company, the models are compatible with the Jev API, which makes testing in applications already built on that interface easier.
In an internal test for domain classification, Clef took 2.2 seconds to fetch, render and classify a site, against 4.7 seconds for a generalist gpt oss 120b model.
Cloudflare prepares custom models for agents
The architecture uses versions of the Qwen family as a base. According to Cloudflare, Clef starts from the Qwen3.8 27B, while Clef flash uses the Qwen3.5 9B, with its own mechanisms to restrict the output to decisions within defined schemas.
The company also introduced a service for fine-tuning with reinforcement learning to adapt Clef to specific workloads. The initiative should evolve into a self-service platform for data capture, training and deployment of custom versions.
The launch reinforces an architecture in which agents can split tasks among different models. Larger LLMs are responsible for generation and open tasks, while specialized systems such as Clef take on fast and structured decisions.
