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 AI post classified by Jev as Open & local models (a tool drop), kept by the AI 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 · 0 likes · 696 views · view the original post on X. Kept by the AI Radar as Open & local models. Tools mentioned: jev-visual.
More AI work like this
- 🚀 Who wants a free virtual RTX Pro 6000? — @ashxhart
- Meet Thomson-1.0-Small-APEX-I-MiniPlus-V2.1-GGUF: a quantized MoE reasoning model in… — @HuggingModels
- Meet T-Search-APEX-I-MiniPlus-V2.1-GGUF: a quantized MoE model built for reasoning and… — @HuggingModels
- Sweeping relaxed acceptance knobs to hit 21.6 TPS DeepSeek v4.1 on M5 tonight.… — @TheDavidTai
- Meet ppt-pythia-1b-appendix-frequent-permuted-seed3407-stage2: a 1B param text generator… — @HuggingModels
- Xiaomi’s MiMo V2.6 Pro is now the #1 global open-source model on Artificial Analysis. — @heyshrutimishra
- Ternary-Bonsai-2-27B-Abliterated has reached 10,000 downloads! Thanks to everyone who's… — @Hikari_07_jp
- Meet Victoria: a text generation model that's lean and mean. It uses MoE pruning, REAP,… — @HuggingModels
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. 44.4k posts from 5k X accounts over the last 14 days, 1.8k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-23 14:14 UTC. Full method.