Cloudflare Launches Clef: Open-Source Decision Models with Vision and RL Fine-Tuning
Cloudflare introduces Clef and Clef-flash, open-source decision models offering high-speed, structured classification for text and images, available both on Workers AI and to run locally. A new reinforcement learning platform for fine-tuning is also released.
What changed?
Cloudflare released Clef and Clef-flash, two open-source decision models trained for fast, accurate, and structured classification. These models are available under the Apache 2.0 license on Hugging Face and are also hosted on Cloudflare Workers AI for edge deployment. Clef introduces a vision encoder, enabling both image and text classification—unlike prior models such as Jev, which only supported text. The models feature a 64k context window (double that of Jev), strictly typed schema outputs with calibrated probabilities, non-autoregressive inference for low latency, and full compatibility with the Jev API. In addition, Cloudflare unveiled a reinforcement learning platform and fine-tuning service, allowing developers to customize Clef for their specific workflows using their own data.
Why does it matter to an everyday developer?
Clef enables developers to add fast, reliable decision-making capabilities to workflows without relying on general-purpose LLMs, which are slower and less predictable for these tasks. Key benefits: - Speed: Clef is significantly faster than large LLMs for classification tasks (e.g., 2.2s versus 4.7s in internal tests). - Structured outputs: Delivers typed, schema-bound results with probabilities, simplifying result handling in code. - API compatibility: Drop-in replacement for Jev in existing applications. - Vision support: Directly classify images as well as text, enabling agentic and multimodal workflows. - Open-source accessibility: Run locally, audit code, or deploy to custom environments. - Fine-tuning: Tailor the model to domain-specific needs using your labeled data, with support for reinforcement learning-based optimization. These capabilities help streamline automation in support, security, content moderation, routing, and any application needing high-confidence, deterministic decisions at scale.
