Cloudflare has released two new AI models, Clef and Clef-flash, that are built not to write text but to make fast, structured decisions. Hosted on Workers AI and fully open-sourced on Hugging Face under an Apache 2.0 license, the models slot into a growing category called “decision models,” which promise cheaper and more consistent outputs than general-purpose large language models (LLMs) for tasks like classification and routing. Alongside the models, Cloudflare is also launching a reinforcement learning (RL) fine-tuning product so customers can adapt Clef to their own use cases.

The release matters because it tackles a real bottleneck in agentic AI systems: making quick, reliable decisions without the cost, latency, or unpredictability of a full LLM call. For developers building automated workflows, that could mean faster pipelines and lower compute bills.

What is a decision model, and how is it different from an LLM

Decision models are related to classifier models, which have existed for years, but are designed to be more flexible. According to Cloudflare, the concept was popularized recently by Typesafe AI’s Jev System One model. A decision model takes in some input, such as a piece of text or an image, and returns a bounded, structured output with probabilities attached rather than freeform generated text.

For example, a decision model can take a customer support message and return a typed answer indicating whether it is urgent and which team should handle it, each with a confidence score. Code downstream can then use that score to route the ticket, escalate it, or hand it to a human. Importantly, Cloudflare says these models can work across new categories without needing to be retrained every time a new classification type is introduced.

This is a deliberate contrast with LLMs. LLMs are open-ended and good at reasoning, generating text, and calling tools, but they are largely non-deterministic and can be slow and expensive for a task as simple as “is this urgent or not.” Decision models trade that open-endedness for speed, consistency, and lower cost on narrower, structured tasks.

How Clef performs in practice

Cloudflare says Clef is currently leading the Jev Decision Index, a public benchmark used to compare decision models, and that full results are viewable on a live benchmark demo site. The company also shared an internal example from its own Threat Intelligence team, which uses Clef to classify website domains.

In that workflow, Clef is given a domain (fetched and rendered via Browser Rendering) and returns multiple weighted classifications at once, for instance a 95% chance a site is a fashion website, 85% ecommerce, and less than 1% phishing. Cloudflare reports that this full fetch-render-classify process took Clef 2.2 seconds, compared with 4.7 seconds for its fastest general-purpose LLM, gpt-oss-120b, in the same workflow, while gpt-oss-120b only returned two classifications instead of several. Cloudflare frames this as roughly a 2x latency improvement with richer output, which it says can meaningfully speed up identification of malicious or legitimate domains at scale.

Why Cloudflare named it Clef

The name is a deliberate musical reference. In music notation, a clef is the symbol at the start of a staff that assigns pitch names to the lines and spaces, effectively setting the context for everything that follows. Cloudflare says it chose the name because a decision model similarly defines the context of a situation and the actions (“notes”) that follow from it, with the added bonus that “CF” nods to the company’s own initials.

What sets Clef apart from other decision models

Cloudflare acknowledges the decision model space is getting crowded, but points to a few differentiators for Clef:

• Vision support: Clef includes a vision encoder, letting it classify images and visual content directly. Cloudflare notes that Jev, by comparison, currently handles text classification only.
• Larger context window: Clef supports a 64k token context window, double Jev’s 32k, allowing more input state to be fed into a single classification call.
• Competitive accuracy: Cloudflare says Clef scores competitively against other decision models on the market, citing its benchmark leadership on the Jev Decision Index as evidence.

Two versions are available: Clef, the full model, and Clef-flash, presumably a lighter, faster variant for latency-sensitive use cases, though the source does not detail the specific tradeoffs between the two beyond naming both as available on Workers AI.

What this means for developers

Because both models are Jev-API compatible, developers already building against Jev’s decision model interface should be able to swap in Clef or Clef-flash with minimal integration work, either through Cloudflare’s hosted Workers AI endpoints or by downloading the open weights from Hugging Face to run locally.

The new RL fine-tuning product is arguably the more significant piece for teams with specialized needs: it lets customers adapt Clef to their own classification categories and domain-specific data rather than relying solely on the general-purpose base model. This matters for any team whose “urgent vs. not urgent,” “phishing vs. legitimate,” or similar classification problem doesn’t map cleanly onto a generic benchmark.

In practice, the pitch is that teams building agentic systems, where software needs to autonomously gather context, decide, and act, can use a decision model like Clef as the fast, cheap “decision layer” in a pipeline, reserving slower and costlier LLM calls for the genuinely open-ended reasoning steps. Cloudflare suggests this pattern generalizes well beyond threat intelligence to any workflow requiring quick, programmatic decisions.

One caveat worth noting: the source material is a product announcement from Cloudflare itself, so performance comparisons (like the 2.2s vs 4.7s figure) come from Cloudflare’s own internal testing rather than independent third-party benchmarking, though the Jev Decision Index results are described as viewable on a public demo site.

Frequently asked questions

Is Clef free to use? The models are open-sourced on Hugging Face under an Apache 2.0 license, meaning developers can download and run them locally without licensing fees. Hosted usage through Workers AI would follow Cloudflare’s standard platform pricing, though specific pricing details are not included in the source material.

Does Clef replace LLMs in agentic workflows? No. Cloudflare positions Clef as a complement to LLMs, handling the fast, bounded decision-making steps, while LLMs remain better suited to open-ended reasoning, text generation, and tool calling within the same workflow.

Can I customize Clef for my own classification categories? Yes, that is the stated purpose of Cloudflare’s new reinforcement learning fine-tuning product, which allows customers to adapt Clef to their specific use cases.

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