Cloudflare 发布 Jev 兼容的决策模型 Clef

Marktechpost AI @Marktechpost原推发布于 2026.10.02

在 X 上打开原推

Cloudflare just released Clef and Clef-flash, its first in-house trained models: open-weight "decision models" that return typed probabilities instead of text. Clef is a 27B model built on Qwen3.8-27B. Clef-flash is a 9B model built on Qwen3.5-9B. Both ship under Apache 2.0. The core innovation is a prefill-only, non-autoregressive decision step. The frozen Qwen backbone runs 1 forward pass, then a small joint schema head routes evidence to each question and scores every option jointly. No tokens are generated, so there is nothing to parse. Training pairs rank-256 LoRA adapters with label-smoothed cross-entropy and a Brier loss for calibration, plus RLCD as a secondary objective. On Cloudflare's internal Decision Index run, Clef-flash posts a 38.8 ms median latency versus 524.1 ms for TypeSafe's Jev, and Clef scores 94.20 on BANKING77 versus Jev's 79.74. It is a strategic trade-off though: Jev still leads on knowledge-heavy tests like GPQA Diamond (78.3 vs 48.0) and MMLU-Pro (82.7 vs 65.9). With image and video input, a 64K context window and a Jev-compatible System One API, it is optimized for routing, triage and guardrails inside agentic workflows...... Full analysis: https://www.marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/ Model: https://huggingface.co/Cloudflare/clef Docs: https://developers.cloudflare.com/workers-ai/models/clef/ @Cloudflare @michellechen