Cloudflare Releases Clef and Clef-Flash Decision Models
Cloudflare launches Clef and Clef-flash open-weight decision models on Workers AI, offering typed probabilities and Hugging Face self-hosting.
Stock photo for illustration only, not from the actual event
- Cloudflare releases Clef and Clef-flash open-weight decision models
- Trained from Qwen3.8-27B and Qwen3.5-9B while keeping vision encoders
- Available instantly through Workers AI, REST API, and AI Gateway
- Scored highest on 7 out of 10 benchmarks in Cloudflare's Decision Index
Cloudflare has officially released two new open-weight decision models named Clef and Clef-flash. Unlike traditional large language models that generate tokens one by one and require downstream parsing, these decision models are built to answer a fixed set of questions about a given input. Both models are currently live on Workers AI, and their weights are made available on Hugging Face for self-hosting setups.
A decision model restricts its output to a predefined structure. Specifically, Clef supports three distinct question types, and a single request on Workers AI can handle up to 64 questions alongside a maximum of 4 images.
Stock photo for illustration only, not from the actual event
Under the hood, Clef is post-trained from Qwen3.8-27B, whereas Clef-flash originates from Qwen3.5-9B, with both architectures retaining their original vision encoders. The inference process executes in two sequential stages:
- The backbone runs a single prefill-only pass across the input state and questions.
- A small transformer component, known as the joint schema head, processes the final hidden states to route evidence and score all choices jointly.
During training, the development team froze both backbone models and jointly optimized the routing head using rank-256 low-rank adapters. The loss function pairs label-smoothed cross-entropy with a Brier loss for calibration, while a secondary objective called Reinforcement Learning for Calibrated Decisions (RLCD) assigns partial credit to adjacent ordinal choices.
Cloudflare's strategic shift toward decision models that return typed probabilities rather than unstructured text addresses a major bottleneck in enterprise AI integration. By removing the need for complex string parsing, these models streamline automated pipelines like invoice processing and threat detection. Deploying them directly on Workers AI bridges the gap between high-performance research models and production-grade developer infrastructure, lowering the friction for building reliable AI applications.
According to Cloudflare's evaluation on the Decision Index 0.2.1 suite, Clef achieved top scores in 7 out of 10 shortlisted benchmarks. Nevertheless, TypeSafe AI's Jev model—launched on September 15, 2026—maintains leads in several general academic metrics, including GPQA Diamond (78.3 versus 48.0), MMLU-Pro (82.7 versus 65.9), and BBH (92.9 versus 73.7). In TypeSafe's internal workflow evaluations, however, Clef slightly outperformed Jev in invoice processing (64.7 to 61.8) and customer service tasks (76.3 to 76.0).
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment