让小模型看一张图回答一串选择题，常见做法是让它输出 JSON，可问题一多格式就开始出错，代码那头直接解析失败。
This is a dev post classified by Jev as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
让小模型看一张图回答一串选择题，常见做法是让它输出 JSON，可问题一多格式就开始出错，代码那头直接解析失败。 jev-visual 换了个思路，参考 TypeSafe 家 Jev 的做法，不让模型生成文字，而是给每个候选答案直接打分，挑分最高的那个。 它是个在 Apple 芯片 Mac 上跑的学习项目，模型用 0.8B 的 Qwen3.5。同一张图一次能问 1 到 64 个问题，选项题、是非题、分等级打分都行。 图片和上下文只算一遍，各个问题共用。作者在 M4 上测了 64 个问题，每个问题单独算要 37.30 秒，共用之后 2.40 秒。 GitHub：http://github.com/hr98w/jev-visual 仓库里带了三个小游戏演示，传送带分拣、打砖块、用摄像头手势控制粒子。 打砖块那个挺实在，小模型盯不住球，作者干脆把画面分成五块，只问球在哪一块。 首次运行要下载大约 600 MB 的模型，之后全在本地跑，有网页界面、命令行和接口三种用法，还配了中文说明。 想弄明白视觉模型推理每一步在干什么，照着 README 给的阅读顺序看代码，会顺很多。
Posted by GitHubDaily (85.3k followers) 1 h ago · 1 likes · 934 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: jev-visual.
More dev work like this
- JevBench v1.4 - now >70 models! — @airesearch12
- 🚨 A new Stealth model "Space bunny" is free on OpenCode for the next week — @LuminaBench
- Finally, an OpenRouter for agent harnesses! — @0x_rody
- STOP making your agent learn only from the chat — @techNmak
- https://v-prism.ai.studio — @athrix_codes
- POV: your boss asks you to build something by next week. You've never built software… — @JJEnglert
- I've built a live feed of our customers' problems to prioritize our roadmap. — @yz_aisdr
- Hice un experimento loquísimo: Una web que se desarrolla a si misma con Inteligencia… — @midudev
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. 22.6k posts from 5k X accounts over the last 21 days, 2.6k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-23 14:48 UTC. Full method.