Introducing Clef: Open-Source Decision Models & RL Fine-Tuning

Published October 1, 2026 | By Michelle Chen, Alex Reneau, and Kevin Flansburg

Recently, the AI community has seen a surge of interest in "decision models," exemplified by the Jev System One from Typesafe AI. While we have had classifiers for years, Jev pioneered a specific concept: a model designed to produce bounded, structured outputs with high speed, low cost, and unwavering consistency. These models are ideal for integration into automated workflows where a definitive choice is required.

Unlike traditional Large Language Models (LLMs), which are often non-deterministic and designed for open-ended reasoning or text generation, decision models are built for precision. Furthermore, they can handle diverse inputs without requiring constant retraining for every new category.

🚀 The Big Reveal: Clef and Clef-flash

Cloudflare is officially launching two proprietary-trained decision models: Clef and Clef-flash. Both are now available via Workers AI.

Key Highlights:

• Performance: Clef currently holds the top spot on the Jev Decision Index.
• Compatibility: Fully compatible with the Jev-API for seamless experimentation.
• Open Source: Available on Hugging Face under the Apache 2.0 license for local deployment.
• Customization: We are debuting a new Reinforcement Learning (RL) platform, enabling users to fine-tune Clef for specific organizational needs.

What exactly is a "Decision Model"?

A decision model acts as the "brain" for an agent, assigning probabilities to various classifications to determine the next course of action.

Example Workflow:

1. Input: A customer support email.
2. Process: The model evaluates urgency and department.
3. Output: A typed response with associated probabilities.

This architecture means that humans must be involved in every agentic decision $\rightarrow$ agents can now autonomously gather context, decide, and execute, only deferring to humans when the probability threshold isn't met.

---

🛠️ Real-World Application: Threat Intelligence

Cloudflare has already deployed Clef within our Threat Intelligence team to categorize website domains. Using Browser Run, Clef analyzes a domain and assigns categories based on probability:

• Fashion Website: $P = 0.95$
• E-commerce: $P = 0.85$
• Phishing: $P = 0.01$

Performance Comparison

By achieving a 2x reduction in latency and higher data density, we can identify malicious domains significantly faster.

---

🎼 Why the name "Clef"?

In music, a clef is a symbol that defines the pitch of the lines and spaces on a staff. Similarly, our decision model defines the domain of context and the subsequent "notes" (actions) that follow. Additionally, the "CF" in Clef is a nod to Cloudflare.

What makes Clef unique?

While the market for decision models is growing, Clef offers three distinct advantages:

• Multimodal Capabilities: Includes a vision encoder, allowing it to classify images (unlike Jev, which is text-only).
• Expanded Memory: Features a 64k context window, doubling Jev's 32k limit for more complex input states.
• High Precision: Competitive scoring across all major quality benchmarks.

---

📊 Benchmarks & Evaluation

We tested Clef against the Jev Decision Index and other popular models. The results demonstrate Clef's superiority in structured decision-making.

Jev Decision Index Results

Typesafe Eval Suite Comparison

Clef outperformed Jev in 3 out of 4 key operational areas:

Latency Analysis

Across 43 benchmarks, Clef models dominated in speed (with the exception of Laya, which is faster but less accurate).

{
  "latency_metrics": {
    "Clef": { "median": "209.3ms", "p95": "238.6ms" },
    "Clef-flash": { "median": "38.8ms", "p95": "122.4ms" },
    "Jev": { "median": "524.1ms", "p95": "536.0ms" },
    "DiffusionGemma": { "median": "84.4ms", "p95": "211.0ms" },
    "Jev Kev-9B": { "median": "51.4ms", "p95": "N/A" },
    "Laya": { "median": "5.8ms", "p95": "N/A" }
  }
}

Figure 1: Visual representation of latency gains across decision models.