Cloudflare Clef: Open-Source Decision Models for AI Agents

209.3 milliseconds. That’s the median time it takes Cloudflare’s new Clef model to make a decision — roughly two and a half times faster than the strongest known competitor. On October 1, 2026, Cloudflare introduced Clef and Clef-flash, two models that don’t write, they decide.

For AI agents that need to choose between yes and no, between multiple options, or along a rating scale in a split second, this is a new building block — and, according to Cloudflare, the first of its kind with openly available model weights. Cloudflare Clef positions itself not as another chat model, but as a specialized infrastructure component for agent workflows that need to make fast, structured decisions instead of producing long text.

Key Takeaways

• Cloudflare Clef (27 billion parameters) and Clef-flash (9 billion parameters) launched as open-source weights under the Apache 2.0 license on Hugging Face on October 1, 2026.
• Both models are hosted on Cloudflare Workers AI under the identifiers @cf/cloudflare/clef and @cf/cloudflare/clef-flash.
• Instead of free text, the models return probabilities for typed answers (yes/no, choice, score), handling up to 64 questions per request.
• In Cloudflare’s own benchmark, Clef is roughly 2.5x faster at the median than competitor Jev, while Clef-flash is about 13x faster.
• Clef supports a 64,000-token context window and additionally processes images and video — a vision encoder that Jev lacks.

What Is Cloudflare Clef? Decision Models Explained

Most well-known AI models are trained to produce coherent text — answers, summaries, code. Cloudflare Clef takes a different approach. The model reads a so-called state, a description of the current situation, together with one or more typed questions. Instead of a text answer, Clef returns a probability for every possible answer.

Cloudflare distinguishes three question types:

• noul, classic yes/no decisions
• choice, selecting one option from a predefined set
• score, rating against a fixed rubric, such as a point scale

According to Cloudflare, a single call can process up to 64 such questions at once. This makes Cloudflare Clef particularly interesting for agent systems that need to make many small decisions in one workflow step — for example, content moderation, website classification, or workflow-logic control. How important robust control mechanisms are for autonomously acting AI agents is also illustrated by our report on a real AI agent hacking attempt.

“Decision Models don’t read a prompt, they read a decision situation.”

This phrasing from Cloudflare’s announcement captures the difference from classic language models: it’s not about creativity, it’s about structured, verifiable decisions.

Cloudflare Clef vs. Clef-flash: Technical Specs at a Glance

Both models are built on existing Qwen architectures but were independently further trained by Cloudflare for the decision use case. The table below summarizes the key technical facts:

The key difference lies not just in model size, but also in functionality: only Clef can process images and video in addition to text. For use cases where pure speed is critical and no visual input is needed, Clef-flash is likely the more practical choice.

Speed Comparison: Cloudflare Clef vs. Jev and gpt-oss-120b

Latency often matters more for decision workflows than for classic chat applications, because agents frequently need to make multiple decisions in sequence before a user even sees a response. Cloudflare therefore ran its own benchmark across 43 test runs, comparing Clef and Clef-flash against competitor product Jev from TypeSafe AI.

The results at a glance:

That makes Clef roughly 2.5x faster than Jev at the median, with Clef-flash about 13x faster. For high-throughput applications such as real-time moderation or bulk classification, this difference can determine practical usability.

A clear gap also shows up against a large general-purpose language model: when classifying a website domain, Clef combined with Cloudflare’s browser-rendering service took 2.2 seconds according to Cloudflare’s own Threat Intelligence team, compared to 4.7 seconds for gpt-oss-120b on the identical task.

Real-World Example: Phishing Detection with Cloudflare Clef

A concrete use case comes from Cloudflare’s internal Threat Intelligence team. The task: automatically classify a website by the probability that it’s a phishing page, an e-commerce shop, or a fashion website.

The process:

1. The target page is loaded via Cloudflare’s browser-rendering service.
2. The resulting state (screenshot, text content, structure) is passed to Clef.
3. Clef answers typed “choice” questions with probabilities for each category.

This combination took 2.2 seconds — significantly faster than the 4.7 seconds gpt-oss-120b needed for the same classification task. For security teams that need to evaluate large volumes of domains in real time, this speed advantage directly shortens response time to potentially malicious sites.

Open Source, License and Availability

A central feature of Cloudflare Clef is its openness. The model weights for both variants are freely available for download under the Apache 2.0 license on Hugging Face. Developers can therefore run the models outside Cloudflare’s infrastructure as well — on their own hardware or with other cloud providers.

In parallel, Cloudflare offers both models hosted via Workers AI, the platform’s serverless inference service. The model identifiers are:

• @cf/cloudflare/clef
• @cf/cloudflare/clef-flash

This dual strategy — open weights plus a hosted service — sets Cloudflare Clef apart from comparable offerings from other vendors that provide their decision models exclusively as closed APIs.

Reinforcement Learning Fine-Tuning: Adapting Clef to Your Workloads

Alongside the model release, Cloudflare announced a new reinforcement learning fine-tuning platform. The goal is to let customers tailor Clef to their own decision workloads instead of working only with the base model.

The rollout happens in two phases:

• First through a forward-deployed engineering team that supports customers directly with fine-tuning.
• Later as a self-service offering that lets developers handle fine-tuning on their own.

For companies with highly specific decision logic — industry-specific moderation rules or internal risk assessments, for example — this offering could become more important long-term than the base model itself, since it enables adaptation to real business rules.

How Cloudflare Clef Compares to the Competition

Cloudflare Clef isn’t the first solution of its kind on the market. Similar “decision” offerings already exist from other major vendors:

The decisive difference lies in openness: as of now, Cloudflare is the only one of the three vendors listed that makes its model weights freely accessible. That gives developers more transparency into the model’s behavior and more flexibility in deployment, regardless of whether they plan to rely long-term on Cloudflare’s own infrastructure.

Who Should Use Cloudflare Clef?

Cloudflare Clef is particularly well suited for teams that:

• build AI agents that need many small, structured decisions rather than long text answers
• have high latency requirements, such as real-time moderation or security classification
• value open model weights and vendor independence
• already work within the Cloudflare ecosystem, e.g. with Workers AI or Browser Rendering

For simple chat applications or creative text generation, however, Clef is not the right fit — classic language models remain better suited for that.

Frequently Asked Questions About Cloudflare Clef

What distinguishes Cloudflare Clef from normal language models?
Clef doesn’t generate free text; it returns probabilities for predefined, typed answers — yes/no, multiple-choice options, or rubric-based scores.

Is Cloudflare Clef free to use?
The model weights are freely available under Apache 2.0 on Hugging Face. Hosted operation runs through Cloudflare Workers AI.

Can Clef process images?
Yes, Clef supports image and video input in addition to text, up to four images per request. Clef-flash is text-only.

How fast is Clef compared to other decision models?
In Cloudflare’s own benchmark, Clef achieved a median latency of 209.3 ms, Clef-flash 38.8 ms, versus 524.1 ms for competitor Jev.

Conclusion

Cloudflare Clef shows how AI infrastructure for agent systems is evolving: away from general-purpose text generators, toward specialized tools for structured decisions. The combination of low latency, open model weights, and direct integration into Workers AI makes Clef a practical option for developers who want to build fast, traceable decisions into their agents.

Developers who want to try Cloudflare Clef can download the model weights on Hugging Face or get started directly via Workers AI using the model IDs @cf/cloudflare/clef and @cf/cloudflare/clef-flash respectively. Teams with specialized needs should also keep an eye on the announced reinforcement learning fine-tuning platform once it becomes available as a self-service offering.

References

• Cloudflare Blog: “Clef: Open-source decision models for AI agents,” blog.cloudflare.com/clef-decision-models/, October 1, 2026.
• Cloudflare Developers Changelog: “Clef on Workers AI,” developers.cloudflare.com/changelog/post/2026-10-01-clef-workers-ai/, October 1, 2026.