Cloudflare Launches Clef, Its Decision Models for Agents

Cloudflare unveils Clef and Clef-flash, two decision models that return probabilities instead of text. Multimodal and touted as very fast, they are compatible with the Jev API and can be executed immediately on Workers AI, with availability on Hugging Face under the Apache-2.0 license.

Availability, License, and Immediate Competition

Clef and Clef-flash are offered on Hugging Face under the Apache-2.0 license and run on Cloudflare's Workers AI platform. The publisher maintains the Clef API compatible with Jev's to simplify any potential transition for clients. These launches come amid active competition: OpenAI introduced a decision API based on GPT-6 Luna at the end of September, accepting both text and images. This approach has been popularized by TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, with Jev positioned as "hallucination-free," meaning it is restricted to a predefined response space without excluding erroneous choices. More broadly, Cloudflare operates from a foundation of network infrastructure, DNS, and security, following recent announcements regarding access control for AI bots, including a measure in July that allows operators to block or allow these bots based on their purpose.

Internal Use Cases and Claimed Human-Free Agents

Cloudflare claims that human intervention is no longer necessarily required in decisions made by agents. The company presents agents capable of collecting contextual elements, making decisions, and acting, while still consulting a human if needed. Internally, Cloudflare intends to use Clef to analyze abuse reports, organize support requests, and distinguish between beneficial and harmful bots. When processing a support stream, Clef can estimate the urgency level of a message and determine the appropriate team, with this information then being utilized by code to route a ticket, trigger an escalation procedure, or assign the case to an operator.

Announced Performance and Comparison to Jev

Cloudflare reports a median latency of 39 milliseconds for Clef-flash and 209 milliseconds for Clef, claiming a speed advantage over competing models. In the provided figures, Jev shows a median latency of just over 524 milliseconds. Both models run on Cloudflare's infrastructure, which claims a benefit from the proximity of edge data centers. According to the publisher's internal measurements, Clef would provide the best decision quality while Clef-flash approaches Jev's accuracy with significantly lower latency.

Multimodal Functions and Web Classification Example

Based on Qwen and compatible with both text and images, these models respond to various predetermined questions for the same data, assigning a probability to each option. Cloudflare's threat intelligence team is already using Clef for website classification: in one presented case, a domain receives a 95% probability of being related to fashion and an 85% probability of being an online store, while the phishing risk is assessed at less than 1%. This sequence of extraction, display, and classification took 2.2 seconds, whereas Cloudflare's most performant general language model required 4.7 seconds and only proposed two classes. Clef processes images where Jev is limited to text and offers a context window of 64,000 tokens, about twice that of Jev. In internal benchmarks, Clef ranks highest on the Jev decision index. More broadly, these models position themselves between LLMs, which are slower and variable but capable of tooling, and traditional classifiers, which are fast but require retraining for each new category.

Architecture, RLCD Training, and Client Customization

According to Cloudflare, Clef is based on Qwen3.8-27B and Clef-flash on Qwen3.5-9B. The base models remain unchanged during training, with additional components trained on synthetic data to derive response options and probabilities from internal calculations. The training employs a variant of RLCD, also used by TypeSafe for Jev, to answer multiple questions in one call and calibrate probabilities based on the actual frequency of correct answers. On the adoption side, Cloudflare is deploying a reinforcement learning service to adapt Clef to client cases, initially with a dedicated team of engineers and a self-service platform planned for later. Clients should be able to create a dataset via AI Gateway, evaluate it in containers acting as RL sandboxes, and then deploy the refined models on Workers AI using a training component. To execute custom models, Cloudflare indicates it uses technology from Replicate, acquired in late 2025. Clef and Clef-flash, as decision models, return probabilities on options rather than textual responses.