dspark-vision

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.

DSpark local vision

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).

When to use

• User gives a local path (/home/mia/pic2.jpg) or image URL and asks what is in it, colors, setting, text, differences between photos, etc.
• Prefer tools over guessing from filenames.

Tools

1. describe_image(path_or_url, question?) — detailed factual description. Pass the user's question as question so the description focuses correctly.
2. ocr_image(path_or_url) — extract visible text only.
3. compare_images(paths, question) — up to 4 images.

Workflow

1. Call describe_image (or OCR/compare) with the absolute path or URL.
2. Read the returned text (or the Error: … message if the file/sidecar failed).
3. Reason over that text yourself on 0731 — including high/max thinking. Do not switch to qwen3-vl-sidecar for the answer.

Notes

• Sidecar must be running (./start-deepseek-v4-flash-dspark.sh, ENABLE_VL_SIDECAR=1).
• Max 4 images per compare call. Huge images are auto-downscaled.
• Do not send image_url to :8888 — 0731 is text-only.