opencode上新了模型,我试了试MiMo和MiniMax M3都可以读图,一开始opencode唯一的缺陷也搞定了,下午发给它grammar friends的前两页,可以支持从图片到md教案的生成。

看一下笔记的长截图效果如下↓
这个太赞了,相当于给opencode加上了眼睛,我随手试了试给它看一看娃画的小兔子,看看它什么反馈。
还支持批改娃的作业——



批改得很细致,该夸的地方夸到细节,该改进的,比如Charlotte's Web书名首字母的大小写、介词搭配等等小问题建议非常详细到位,还要报什么课外班呢?还有opencode非常执着要把批改的内容存到我对应的文件夹里,太顺手和智能了。以下是opencode自己写的读图能力清单,大家可以逐步试玩起来。
🌿 🍃 🌱
读图能力清单 ·Sisyphus
模型验证过「能看图」之后,实际能做的 vs 不能做的
能稳定做有条件能做不能做
| 场景 | 类别 | 我能做什么 | 限制 / 前提 | 状态 |
|---|---|---|---|---|
| 📚 课本/作业批改 | 教育 | 孩子英文/语文手写作业,能识别字迹、批改语法错、给出修改建议和理由 | 字迹不能太潦草,照片不要反光 | 能 |
| 🔢 错题讲解 | 教育 | 数学算式、看图应用题、自然课本截图,给出解题思路 | 需要图片清晰,文字无遮挡 | 能 |
| ✍️ 看图写作 | 教育 | 看一张图,出小作文、看图说话、故事接龙 | 无 | 能 |
| 📖 名著/章节答疑 | 教育 | 拍一页书(Charlotte's Web、Magic Tree House),问"这段什么意思" | 页面要拍正,光线均匀 | 能 |
| 🎨 孩子画作讲故事 | 创意 | 识别画里的元素(动物、场景、颜色),扩展成完整小故事 | 无 | 能 |
| 🖼️ 海报/Logo 评价 | 创意 | 描述画面、评价设计、给改进方向 | 需要相对清晰的图像 | 能 |
| 🎭 名画/艺术品介绍 | 创意 | 识别流派、画派、可能作者、时代背景 | 我不专业,复杂作品可能说错 | 有条件 |
| 😂 梗图/表情包解释 | 创意 | 看图说梗、判断笑点、解释为什么好笑 | 依赖文化背景,海外梗可能认不出 | 能 |
| 🍜 食物/菜谱识别 | 生活 | 看一张菜,告诉你大概是什么、可能做法、热量估算 | 家常菜准,复杂菜系会模糊 | 有条件 |
| 🛋️ 装修/家具参考 | 生活 | 描述风格(北欧/日式/工业),给搭配建议 | 只能描述,不能"生成"或"画" | 有条件 |
| 📍 实物照片问答 | 生活 | "这是什么植物""这包药怎么吃""这地方在哪" | 模糊/小图会误判 | 能 |
| 📄 截图 OCR 提取文字 | 工具 | 把截图里的文字转出来,做成 markdown 表格/清单 | 手写字比印刷字容易出错 | 能 |
| 📊 图表/数据图分析 | 工具 | 读柱状图、折线图、饼图,告诉你"哪一项最大"等结论 | 模糊或非常规图表会读错 | 有条件 |
| 💻 代码截图 review | 工具 | 贴报错截图、代码截图,定位问题、给改法 | 最好能直接给文字代码 | 能 |
| 🖥️ UI 截图提建议 | 工具 | 网页/App 截图,评价布局、配色、可用性 | 只能评价,不能直接改 | 能 |
🚫 确认不能做的(重要)
- 不能直接打开你电脑上的图片文件(除非你先贴/拖给我)
- 不能看视频/动图的动态部分(只能看静态截图/首帧)
- 不能修图、画图、生成图片(只能描述)
- 不能直接操作你的软件(不能"点这个按钮")
- 看不清时不要相信我的猜测——我会直说"看不清"而不是硬猜
📐 怎么用效果最好
1. 图片要清晰
- 光线均匀,不要逆光,不要反光
- 文字部分拍正,不要斜着拍
- 需要批改的字迹,建议
200万像素以上+ 自然光
2. 一次只发 1-2 张
- 多张图混着发容易丢上下文
- 如果是一整页课本/作业,一页一张图
3. 配合文字描述背景
单独发一张孩子的英文作业,vs.说"这是我家孩子写的 StoryFun 作业,5 岁,能拼 helicopter,麻烦看下语法有没有错"→ 后者我能给更针对性的批改。
🎁 真实可用组合(结合你家场景)
- ✅ 孩子 StoryFun 作业拍照 → 批改 + 夸奖(你已用过)
- ✅ Grammar Friends 课本页 → 出小教案(你已用过)
- ✅ 孩子画作 → 描述画面 + 扩写成小故事(试过)
- ✅ Magic Tree House / Charlotte's Web 任一页 → 解释生词 + 提 1 个讨论问题
- ✅ 名著读不懂的段落拍照 → 翻译 + 解释
- ✅ 某次考试的错题拍照 → 整理成"错题本" markdown 存到 Obsidian
大家可以按照清单研究试用有眼睛的opencode,我觉得批改作业是最实用的,直接让opencode把错题笔记整理到obsidian指定文件夹里。感谢您阅读到这里,希望对您也有一点点帮助💗
opencode released new models. I tried MiMo and MiniMax M3, and both can read images. The only flaw opencode had from the start is now fixed. In the afternoon I sent it the first two pages of Grammar Friends, and it can support generating an MD lesson plan from images.
Take a look at the long screenshot effect of the notes below↓ This is amazing. It's basically like giving opencode eyes. I casually tried showing it a picture of the little rabbit my kid drew, just to see what feedback it would give. It can also grade my kid's homework — and it grades in great detail, praising specific details where praise is due, and giving very thorough, targeted suggestions for things to improve, like the capitalization of the first letters in the book title Charlotte's Web, preposition collocations, and other small issues. Why would you even need to sign up for extracurricular classes? Also, opencode is very insistent on saving the graded content to my corresponding folders, which is super convenient and smart. Below is a list of image-reading capabilities written by opencode itself. Everyone can try it out step by step.









🌿 🍃 🌱
Image-reading capability list · Sisyphus
After verifying that the models "can see images," what they can actually do vs. can't do
Can do reliably / Can do under certain conditions / Can't do
🚫 Confirmed things it can't do (important)
It cannot directly open image files on your computer (unless you paste or drag them to me first)
It cannot watch the dynamic parts of videos/GIFs (can only view static screenshots/first frames)
It cannot edit, draw, or generate images (it can only describe)
It cannot directly operate your software (it can't "click that button")
When the image is unclear, don't trust my guesses — I'll plainly say "I can't see clearly" rather than guessing blindly
📐 How to get the best results
1. The image should be clear
Even lighting, avoid backlighting, avoid glare
Shoot the text straight on, don't shoot at an angle
For handwriting that needs grading, it's recommended to use 2+ megapixels and natural light
2. Send only 1-2 images at a time
Sending multiple images mixed together can easily lose context
If it's a full page of a textbook/homework, use one image per page
3. Pair it with a text description of the context
Sending a kid's English homework by itself, vs. saying "This is my kid's StoryFun homework, age 5, can spell 'helicopter,' could you check the grammar for errors?" → with the latter, I can give more targeted feedback.
🎁 Real usable combinations (for your family's scenarios)
✅ Photograph your kid's StoryFun homework → grade it + give praise (you've already used this)
✅ Grammar Friends textbook page → generate a mini lesson plan (you've already used this)
✅ Your kid's drawing → describe the picture + expand it into a short story (tried this)
✅ Any page of Magic Tree House / Charlotte's Web → explain new words + ask 1 discussion question
✅ Photograph a passage of a classic you can't understand → translate + explain
✅ Photograph wrong answers from a certain exam → organize them into a "mistake notebook" markdown and save it to Obsidian
Everyone can follow this list and try out opencode with eyes. I think grading homework is the most practical use — just have opencode organize the wrong-answer notes into the designated Obsidian folder. Thank you for reading this far. I hope it's a little helpful to you too💗