菜单
Local vision for DeepSeek-V4-Flash-0731 via Qwen3-VL sidecar MCP tools. Use when the user attaches or mentions an image path/URL and needs visual facts, OCR, or image comparison — stay on 0731; do not switch to a VL model.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark --skill dspark-vision命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | dspark-vision |
| description | Local vision for DeepSeek-V4-Flash-0731 via Qwen3-VL sidecar MCP tools. Use when the user attaches or mentions an image path/URL and needs visual facts, OCR, or image comparison — stay on 0731; do not switch to a VL model. |
The default model (deepseek-v4-flash-0731) is text-only. Vision comes from
MCP tools that call the on-prem Qwen3-VL-4B sidecar (http://127.0.0.1:8889,
TP=2 across both Sparks).
/home/mia/pic2.jpg) or image URL and asks what is
in it, colors, setting, text, differences between photos, etc.describe_image(path_or_url, question?) — detailed factual description.
Pass the user's question as question so the description focuses correctly.ocr_image(path_or_url) — extract visible text only.compare_images(paths, question) — up to 4 images.describe_image (or OCR/compare) with the absolute path or URL.Error: … message if the file/sidecar failed).qwen3-vl-sidecar for the answer../start-deepseek-v4-flash-dspark.sh, ENABLE_VL_SIDECAR=1).image_url to :8888 — 0731 is text-only.