@hr98w demos a local visual stand-in for Jev: a 4-bit multimodal Qwen-3.5-0.8B maps candidates to fixed labels, samples after prefill, and limits output tokens. It runs on a 16GB M4 MacBook and sorts three-color goods in about 200ms. They say this copies the form of Jev at inference time, not its real method.

Haoran | 公众号：独立开发 (@hr98w) published this on 2026-09-17. @hr98w demos a local visual stand-in for Jev: a 4-bit multimodal Qwen-3.5-0.8B maps candidates to fixed labels, samples after prefill, and limits output tokens. It runs on a 16GB M4 MacBook and sorts three-color goods in about 200ms. They say this copies the form of Jev at inference time, not its real method. 视觉版 Jev 来了 https://t.co/x0kzeE3hdx 仿照着社区的思路，把候选结果映射成固定标签，prefill 完直接采样，限制输出 token，采用多模态的 qwen-3.5-0.8B 4bit 量化部署在本地，16g m4 mbp 顺利运行 叠甲：这只是在 infer 层小小的复现一下 jev 的形式，肯定不如各种成熟推理框架效率高，也肯定不是 jev 的正在原理，但是作为 inference 的小入门仍旧是不错的 demo 如下 200ms 实现《三色货品分拣》 How Jev or TypeSafe is used, in the post's own words: “视觉版 Jev 来了 https://t.co/x0kzeE3hdx 仿照着社区的思路，把候选结果映射成固定标签，prefill 完直接采样，限制输出 token，采用多模态的 qwen-3.5-0.8B 4bit 量化部署在本地，16g m4 mbp 顺利运行 叠甲：这只是在 infer 层小小的复现一下 jev 的形式，肯定不如各种成熟推理框架效率高，也肯定不是 jev 的正在原理，但是作为 inference 的小入门仍旧是不错的 demo 如下 200ms 实现《三色货品分拣》” The post does not name a separate audience. It is for people reading what @hr98w shows here (it links to github.com; it includes media).