Artificial intelligence took another significant step forward today. Cloudflare officially launched Clef and Clef-flash, two new open-source decision models. Consequently, developers can now access fast, consistent, and cost-effective structured outputs.
Cloudflare hosts these new models directly on the Cloudflare Workers AI platform. Therefore, this new class of decision models offers a streamlined alternative to traditional large language models.
What Is a Decision Model?
Large Language Models (LLMs) are inherently non-deterministic, and they often require several seconds to generate text. In contrast, decision models focus on generating typed, structured classifications with exact probabilities.
Furthermore, this architecture delivers several key advantages:
- Autonomous Agent Workflows: Agents can make deterministic, programmatic choices. As a result, systems no longer require constant human intervention.
- Efficient Data Routing: Software can instantly categorize support tickets, route urgent escalations, or identify security threats.
- Flexible Generalization: Models easily handle unseen input categories without requiring constant retraining.
Key Advantages of Clef Over General LLMs
According to internal benchmarks on the Jev Decision Index, the Cloudflare decision model excels across several core metrics:
- Ultra-Low Latency: Cloudflare’s Threat Intelligence team tested Clef on web domain classification. Clef analyzed and classified domains in just 2.2 seconds. Meanwhile, general LLMs like gpt-oss-120b took 4.7 seconds for the exact same workflow.
- Multimodal Capabilities (Vision Encoder): Clef features an integrated vision encoder. Thus, it evaluates visual imagery alongside text to achieve higher classification accuracy.
- API Compatibility: The model fully supports the Jev API spec. Because of this, developers can easily plug Clef into existing codebases.
- Open-Source License: Cloudflare released the model on Hugging Face under the Apache 2.0 license for local deployment.
Reinforcement Learning (RL) Fine-Tuning Platform
Alongside Clef, Cloudflare introduced a new Reinforcement Learning (RL) fine-tuning platform. This tool allows enterprise teams to fine-tune the model on their own domain-specific datasets. As a result, teams can significantly increase classification precision for specialized workflows.
For complete technical documentation, you can visit the official announcement on the Cloudflare Blog.
Why Is It Named “Clef”?
In musical notation, a clef is a symbol at the beginning of a staff that sets the pitch context for subsequent notes. Cloudflare chose the name Clef because this decision model establishes the context before downstream actions execute. Additionally, the initials “CF” pay tribute to Cloudflare.