devwithjev
— reading now— views
Submit a build

Untitled

Jev 居然输给了一个 1.88B 本地模型？我用俄罗斯方块重新测了一遍 当我看到 this-that-model-1.0 在 68 道决策题上做到 94.1% 准确率，而 Jev 是 76.5% 时，我的第一反应是： 这是真的假的？ 一个只有 1.88B 参数、可以在 Mac 本地运行的小模型，真的能在决策任务上超过 Jev 吗？ 所以我做了一个俄罗斯方块，让两个模型使用相同的棋盘规则、随机种子和方块顺序。程序负责计算所有合法落点，模型只负责决定方块应该放在哪里。 1️⃣ 一开始，Jev 的表现更好 最初，我先过滤掉明显更差的落点，再把剩下的多个候选位置同时交给模型。 在这一模式下，Jev 的表现更加稳定。 我观察到的一轮里： Jev 消除了 5 行，this-that 消除了 3 行。 如果只看到这里，很容易得出结论：本地小模型还是不如 Jev。 但我后来意识到，这种问题可能并不是 this-that 最擅长的形式。 2️⃣ 换成“二选一”后，结果反了 this-that 本质上是一个 typed decision model，更擅长候选明确、边界清楚的选择题。 所以我把决策方式改成了淘汰赛： A 和 B，哪个落点更好？ 胜者进入下一轮，继续和其他候选位置比较，直到选出最终落点。 换成这种模式后，我观察到的一轮结果变成了： this-that 消除了 7 行，Jev 只消除了

View on X
阿蔺A-Lin@alin_zone𝕏
Jev 居然输给了一个 1.88B 本地模型？我用俄罗斯方块重新测了一遍 当我看到 this-that-model-1.0 在 68 道决策题上做到 94.1% 准确率，而 Jev 是 76.5% 时，我的第一反应是： 这是真的假的？ 一个只有 1.88B 参数、可以在 Mac 本地运行的小模型，真的能在决策任务上超过 Jev 吗？ 所以我做了一个俄罗斯方块，让两个模型使用相同的棋盘规则、随机种子和方块顺序。程序负责计算所有合法落点，模型只负责决定方块应该放在哪里。 1️⃣ 一开始，Jev 的表现更好 最初，我先过滤掉明显更差的落点，再把剩下的多个候选位置同时交给模型。 在这一模式下，Jev 的表现更加稳定。 我观察到的一轮里： Jev 消除了 5 行，this-that 消除了 3 行。 如果只看到这里，很容易得出结论：本地小模型还是不如 Jev。 但我后来意识到，这种问题可能并不是 this-that 最擅长的形式。 2️⃣ 换成“二选一”后，结果反了 this-that 本质上是一个 typed decision model，更擅长候选明确、边界清楚的选择题。 所以我把决策方式改成了淘汰赛： A 和 B，哪个落点更好？ 胜者进入下一轮，继续和其他候选位置比较，直到选出最终落点。 换成这种模式后，我观察到的一轮结果变成了： this-that 消除了 7 行，Jev 只消除了
Sep 23, 2026X postsView on X

Also filed under Tools & apps

  • Jev: a "System One" model that only makes decisions, $0.042 per million tokens with output free. 40-second explainer of why it matters for automation

    Jev: a "System One" model that only makes decisions, $0.042 per million tokens with output free. 40-second explainer of why it matters for automation Jev launched on Sep 15 from TypeSafe AI (founder Diogo Almeida, ex-OpenAI, credited on ChatGPT/InstructGPT). It does not generate text; it returns typed decisions (Choice / Score / true-false) with confidence. Price is $0.042 per million input tokens, output free. In its first week it became the fastest-adopted model on Vercel's AI Gateway, TypeSafe paused signups on Sep 22, Browser Use released jev-ultrafast (a web agent where Jev picks each action; Google Flights search in 7.1 s), and jaredpalmer/kev is an Apache-2.0 clone built on Qwen3.5 you can run yourself. My take in the video: the expensive part of automation was never the writing, i

  • Untitled

    There's a new AI model that refuses to talk. While everyone debates what a "decision model" is even for, builders already shipped the answer. 7 things people built with Jev in 9 days. Number 5 broke my brain 👇

  • Untitled

    I loved this idea so much that it deserves a share ❤️ I recently asked people to share what they’ve built with JEV, and then this guy 👤 shows up with an entire website dedicated to different JEV use cases! That’s honestly pretty cool. If you’re curious about JEV and want to see how it can actually be applied, definitely check this out. Huge shoutout to Animesh Mishra (@thatcasualvc) for the wonderful share 🙌 This is exactly the kind of stuff I love seeing from builders. 🚀

  • I built a Jev-based model router for Pi for automatic model selection

    I built a Jev-based model router for Pi for automatic model selection When Jev came out, one of my first thoughts was: could something this fast and cheap pick which model should handle an agent’s next turn? So I built plugin for Pi that uses Jev to choose the model and thinking effort. The idea is to send routine work to cheaper models and reserve the expensive ones for harder tasks. In Pi, the model mappings are configurable, and you can pin a model when you disagree with the router. I’ve been using them for a few days, and so far they’ve been working well for me. In my own evaluation, routed usage came to $141 at list prices versus an estimated $166 for Opus 5.5 on every request - about 15% less. That’s one developer’s data, though. The baseline is a pricing comparison, not a separate