Cloudflare Releases Open Decision Model Clef, Adding Image Input To The Jev Formula
Clef turns the decision-model format Jev popularized into something Cloudflare runs in production on its own threat intelligence, and Jev's creator is publicly arguing that the benchmark numbers deciding which model wins do not measure what he built for.
Reporting from 1 source: GIGAZINE.
Cloudflare has released Clef and Clef-flash, open decision models under Apache License 2.0, downloadable from Hugging Face. Unlike Jev, which handles only text, Clef accepts image input. GIGAZINE reports Clef scored above Jev on Decision Index 0.2.1 while running faster, and Cloudflare's threat intelligence team already uses it to classify websites. Diogo Almeida, who made Jev, questioned benchmark validity.
Cloudflare names both variants in the release: the standard Clef and the lighter Clef-flash. The name comes from the clef symbol that sets pitch on a musical score. In Cloudflare's own testing, the threat intelligence team uses Clef to sort websites under review, returning scores like a 95 percent chance a site is a fashion retailer, 85 percent that it is an online store, and under 1 percent that it is a scam. The same classification takes gpt-oss-120b 4.7 seconds and Clef 2.2 seconds. Ollama supports running the Clef series.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.