Cloudflare Releases Clef, Open-Weights Multimodal Decision Models
Cloudflare has open-sourced its first in-house decision models, Clef and Clef-flash, paired with a new reinforcement learning (RL) fine-tuning service. These models process complex inputs like text, JSON, images, and video to evaluate states and output probabilities for typed question schemas. This represents a major infrastructure provider offering specialized, lightweight open-weights decision models rather than general-purpose text generators. It allows developers to build structured, probabilistic decision pipelines directly on edge infrastructure with workflow-specific RL tuning. Clef-flash is a 9-billion parameter multimodal model that analyzes structured or unstructured state inputs against predefined typed schemas. Cloudflare's accompanying RL platform integrates AI Gateway, Containers, and Workers AI to streamline custom fine-tuning for specialized applications.
## BACKGROUND
Open-weights models allow developers to inspect, host, and fine-tune pre-trained neural networks on their own hardware or cloud providers. Unlike traditional text-generation LLMs that generate open-ended responses, decision models evaluate current state data against concrete rules or options to assign probability scores for action selection.