Cloudflare: New Clef model removes the need for human oversight in AI agents

Meta description: Cloudflare releases Clef and Clef-flash, decision models for AI agents that report 39 ms and 209 ms median latencies against Jev across 43 benchmarks.

Tags: Cloudflare Clef, AI agents, decision models, Qwen3.8-27B, Workers AI

Featured image: Abstract technology scene with probability vectors and edge network paths, no real people, no company logos, no text.

Cloudflare Clef and Clef-flash are two decision models for AI agents that return brief classifications with probabilities instead of long text answers. The company pitches latency as the main advantage against the current leader, Jev. In a support case, Clef can judge a customer message's urgency and name the responsible team, while downstream code can route a ticket, trigger escalation, or send the issue to a person.

Background: A routing gap opens between classifiers and agents

Instead of writing a paragraph, a decision model answers several preset questions for one input and attaches a probability to each option. This niche sits between large language models and traditional classifiers. Language models can reason and call tools, yet their outputs vary and they can be slow. Classifiers respond quickly, but each new category needs retraining.

Cloudflare states, "a human does not necessarily need to be in the loop for agentic decisions anymore." Agents can "programmatically gather context, make decisions, and take actions on tasks, or defer to a human when needed." The product name borrows from music notation: a clef assigns pitches to the lines of a staff, and a decision model sets the framework for later actions. The phonetic resemblance to "Jev" is probably no accident.

Development: Jev's lead, OpenAI's entry, and Clef's benchmark pitch

TypeSafe AI popularized the approach; it was founded by former OpenAI researcher Diogo Almeida and introduced Jev in mid-September. TypeSafe markets Jev as a model "without hallucinations," though that claim only guarantees answers stay within predefined options and does not prevent incorrect choices. In late September, OpenAI followed with a Decisions API built on GPT-6 Luna that also accepts context as text or images.

Cloudflare primarily runs a global network for content delivery, DNS, and security services and had attracted little attention for its own AI models. Its recent AI coverage focused on access control: in July, it let site operators block or allow AI bots based on purpose. Now Clef is positioned as a direct answer to Jev, and the API remains fully compatible so customers can switch easily.

Across 43 benchmarks, Cloudflare says Clef and the smaller Clef-flash are faster than all relevant competing decision models. The smaller Clef-flash can return a decision in 39 milliseconds. Reported median latencies are about 39 milliseconds for Clef-flash and about 209 milliseconds for Clef, versus just over 524 milliseconds for Jev. Both run directly on Cloudflare infrastructure, which the company says benefits from proximity to edge data centers. In Cloudflare's self-reported numbers, Clef gives the highest decision quality, while Clef-flash nearly matches Jev's accuracy at a fraction of the latency.

Cloudflare says Clef can also process images, whereas Jev is limited to text so far. Its 64,000-token context window holds twice as much input as Jev's. Clef also leads on the Jev Decision Index in Cloudflare's own benchmarks.

Technical path and early use cases

Clef is based on Qwen3.8-27B and Clef-flash on the smaller Qwen3.5-9B. Cloudflare leaves the base models unchanged during training and uses its own synthetic data to train extra components that derive answer options and probabilities from internal computations. The company uses its own variant of Reinforcement Learning for Calibrated Decisions (RLCD), the method TypeSafe used for Jev. RLCD trains a model to answer several input questions in one call and aims to assign probabilities that match actual correctness rates. Cloudflare had previously experimented with DiffusionGemma to derive fixed decision values from a language model's internal computations.

Cloudflare's threat intelligence team is already testing Clef to classify websites. In one example, it assigns a domain a 95 percent probability of being a fashion website and 85 percent of being an online store, while phishing probability is under one percent. Fetching, rendering, and classifying took 2.2 seconds. The company's fastest general-purpose language model took 4.7 seconds for the same process and returned only two categories.

Latest: Fine-tuning, internal rollout, and open availability

Alongside the release, Cloudflare is rolling out a reinforcement learning service that lets customers tailor Clef to their own tasks. A team of forward deployed engineers will initially handle fine-tuning with customers, with a self-service platform planned for later. Customers can build a dataset by logging requests through AI Gateway, then evaluate those requests in containers that serve as an RL sandbox. A new trainer component will let them deploy the fine-tuned model on Workers AI. To run custom models, Cloudflare uses technology from Replicate, which it acquired in late 2025. The planned self-service platform will log production data, fine-tune Clef with reinforcement learning, and redeploy the customized model directly.

Both models run on Cloudflare's Workers AI platform and are available on Hugging Face under the Apache-2.0 license. Cloudflare says it plans to use Clef internally to review abuse reports, sort support requests, and distinguish useful bots from harmful ones.

The launch puts Cloudflare in the agent-routing phase, where decision models must be fast, calibrated, and adaptable. The current evidence is mostly self-reported benchmarks and internal workflow numbers, while Jev-compatible APIs and open availability lower the barrier for teams testing probability-based routing.