Jev Adds Typed Decision Probabilities but Trails Some Rivals
A Juejin analysis reports that TypeSafe AI released Jev on September 15, 2026. Jev accepts state and typed questions and returns option probabilities for classification tasks without generating free-form text. Version 1.14 added the Box primitive for image localization, while JevBench Capability fell from 77.1 for version 1.13 to 75.0; Calibration remained near 90, declining from 90.6 to 90.1. The analysis says Jev offers low latency but is mid-to-high priced among decision models, with some faster or more accurate competitors. It also notes that type safety confines answers to supplied options but does not guarantee they are correct.
The report illustrates a decision-model interface that separates classification from token generation. It also shows why type safety, calibration, latency, price and benchmark performance need separate evaluation rather than reliance on claims of overall superiority.
2026 年 9 月 15 日,TypeSafe AI 发布 Jev。 六到七周内,Cloudflare、微软、OpenAI、亚马逊全部推出同类;OpenRouter 上一周 15 亿次调用打到 TypeSafe 的模型上;10 月,决策模型占了 Vercel 模型路由四分之一以上的请求。 但同一时间,Jev 在第三方 JevBench 的托管 API 榜上排第 15–18 名;一个 2,000 封邮件的钓鱼判定测试里,它只有 62.6%,输给 Claude Haiku 4.5 的 81.3%。 这两个事实同时为真。这篇文章把它们拼起来看:Jev 到底是什么、快和便宜从哪来、现在有哪些能打的同类、什么场景该用、以及怎么在半天内跑通第一个可用版本。
先说结论:
- Jev 不是「更小的 LLM」,它是一个接口不同的模型:输入
state加一组带类型的问题,输出选项概率,不做 token 生成。 - 快和便宜来自「一次前向 + 所有问题并行求值」,但官方那个「193 倍 / 444 倍」是对着「逐 token 吐 JSON」的稻草人测出来的,自己算一遍更靠谱。
- 它真正值钱的是校准与类型安全,不是「零幻觉」这个营销说法。类型安全保证答案在你的选项里,不保证答案是对的。
- Jev 已经不是最强的那一个,但它的接口正在变成事实标准;选型别只看榜首,落地只信你自己那 300 条标注。
#TypeSafe AI#Jev#Jev 1.13#Jev 1.14#JevBench#Artificial Analysis#OpenAI#Microsoft#Amazon#Cloudflare#Claude Haiku 4.5#Vercel