牛逼！Jev 刚火没两天，已经有人搞了个开源的Jev🤣
This is a AI post classified by Jev as Open & local models (a model release), kept by the AI Radar because it carries real work, not commentary.
牛逼！Jev 刚火没两天，已经有人搞了个开源的Jev🤣 @bespokelabsai 发布了一个开源版的Jev：Bespoke Nimble。 更离谱的是，整个项目是基于 Qwen3.5-9B ，只花了大约两天。 数据、模型权重、训练配方、推理代码全部公开，而且没有从 Jev 蒸馏，只拿 Jev 做评测参照。 最终方法是在 Qwen3.5-9B 上做了一次很轻量的 LoRA微调。 只用了2676 条训练数据，结果在他们自己的 324 条 held-out 测试集上拿到了90.12%，而原版的Jev是93.21%。 一个 9B 模型，已经做到距离 Jev 只差 3 个百分点。 但我觉得最值得看的不是这个分数，而是他们到底怎么训的。 核心方法叫 contrastive data curation。 简单说，他们会构造两个几乎完全一样的样本，只修改一个足以改变最终决策的关键事实，让正确答案发生翻转。 模型因此被迫学会一件事： 到底哪条证据，才真正应该改变决策。 这其实特别符合 Jev 这类模型的用途。 它本来就不是拿来陪你聊天的，而是塞进 Agent 和软件系统里，做路由、审核、打分、策略判断这些高频的小决策。 推理方式上Nimble 不生成 CoT，甚至不生成 JSON。 它直接读取候选答案 token 的 logits，再把它们转成概率，最后由程序拼成
Posted by Max For AI (39.4k followers) 10 h ago · 322 likes · 37.8k views · view the original post on X. Kept by the AI Radar as Open & local models. Tools mentioned: nimble, bespoke-nimble-9b.
More AI work like this
- dear rtx 3060 owners, and every 12gb card behind it. bonsai 2 27b dense went from 26 to… — @sudoingX
- Nimble is able to process images! — @madiator
- Woah...it claims to be open Jev. And it's 100% opensource. — @Saboo_Shubham_
- your rtx 3060 was running bonsai 2 at 26 tok/s this morning and now does 40 tok/s, i… — @sudoingX
- Meet DeepSeek-V4.1-Flash: a blazing fast image-text-to-text model. It takes both images… — @HuggingModels
- Meet Edge0-35B-A3B-preview: a 35B parameter MoE model built for edge inference. It runs… — @HuggingModels
- Putting Jeff-1 out into the world, this is an attempt at a Jev like model, built on qwen… — @DJLougen
- Train a 9M parameter language model from scratch in five minutes. — @tom_doerr
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 27.2k posts from 4.8k X accounts over the last 14 days, 1.2k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-19 18:45 UTC. Full method.