Cloudflare 开源 Clef，对标 Typesafe 的 Jev
Cloudflare 发布开源决策模型 Clef，宣称 API 与 Jev 兼容，在延迟和 Decision Index 上压过 Jev，但 Jev 在 When2Call、BRIGHT 和 agent 追踪工作流上仍占优。
Cloudflare released Clef, an open-source decision model, and put it straight up against Typesafe AI's Jev. Quick context: a decision model doesn't write text like an LLM. You give it questions (which team? is it urgent?) and it returns typed answers with probabilities your code can branch on. Think alert triage and ticket routing. Clef vs Jev, from Cloudflare's launch post: - API: Cloudflare says Clef is fully Jev-compatible, so trying it shouldn't need a rewrite - Input: Clef takes text and images, Jev text only - Context window: 64k vs 32k - Median latency: 209 ms vs 524 ms, and Clef-flash at 39 ms - Decision Index: 61.2 for Clef vs 57.9 for Jev - Weights: Clef is open on Hugging Face under Apache 2.0, Jev is closed Where Jev still wins in the same post: When2Call, BRIGHT, and the agent-trace observability workflow. Clef's scores are self-reported, on benchmarks Cloudflare chose. If you already run Jev, compare both on your own data before switching. Images: Cloudflare's launch post.