Cloudflare releases Clef and Clef-flash decision models on Workers AI

Clef: Clef is a Cloudflare-developed decision model with a vision encoder, 64k context window, and architecture optimized for producing bounded structured outputs with calibrated probabilities. It serves as the primary precision model in the company's new family of decision models and leads on relevant benchmarks. The model is hosted on Workers AI, fully open-sourced on Hugging Face, and designed for seamless integration into workflows requiring fast classifications such as threat intelligence. Clef-flash: Clef-flash is the lighter, latency-optimized variant in Cloudflare's Clef family of decision models, built on a smaller Qwen backbone. It prioritizes speed for time-critical decisions while maintaining competitive accuracy and full API compatibility with existing decision model standards. Like its larger counterpart, it is hosted on Workers AI and open-sourced for local use or experimentation. Cloudflare: Cloudflare provides cloud infrastructure, security services, and AI platforms including Workers AI for edge-based model inference. In this news, the company is announcing the release of its proprietary decision models Clef and Clef-flash along with a new RL fine-tuning service. These offerings build on Cloudflare's internal experiments adapting base models for deterministic classification tasks in agentic systems. Workers AI: Workers AI is Cloudflare's edge AI inference platform that runs models close to users for minimal network latency. It hosts the new Clef models and supports Bring Your Own Model deployments along with upcoming fine-tuning capabilities. The platform enables combining decision models with other AI workloads in agentic pipelines directly on Cloudflare infrastructure. Hugging Face: Hugging Face is the leading platform for sharing, hosting, and collaborating on open-source machine learning models. Cloudflare chose it to release the weights and code for Clef and Clef-flash under an Apache 2.0 license. This allows developers to download, run locally, or further experiment with the models outside of Cloudflare's hosted environment. AI Model Types: Decision models deliver deterministic, schema-bound classifications that integrate directly into agent code, unlike the open-ended text generation typical of large language models. Edge AI Deployment: Running models on edge infrastructure supports low-latency inference suitable for real-time decision points in automated workflows. Open-Source Release: Releasing models under permissive licenses on established repositories enables community access, local execution, and custom experimentation beyond hosted services.