Image Recognition with DeepSeek

DeepSeek's API is text-only. Scholardo bridges the gap by running a local proxy that uses Zhipu GLM-4.6V to describe images, then forwards the text to DeepSeek — so your DeepSeek sessions can "see" images without any local GPU load.

Why Zhipu GLM-4.6V

• Reachable inside China without a VPN
• OpenAI-compatible endpoint — Scholardo's vision proxy needs no changes
• Free tier (glm-4.6v-flash) is sufficient for image-to-text transcription
• Cloud inference — no OOM issues from running a local VLM

Step 1 — Get a Zhipu API Key

1. Go to open.bigmodel.cn and log in.
2. Open API Keys and create a new key (format: xxxxxxxx.xxxxxxxx).

Step 2 — Add the provider in Scholardo

Open Settings → AI API, scroll to the OpenAI-compatible (custom provider) section, and add:

Field	Value
Name	Zhipu GLM (or anything)
Base URL	https://open.bigmodel.cn/api/paas/v4
API Key	your Zhipu key

WARNING

Enter the URL up to .../paas/v4 — do not append /chat/completions. The proxy adds that itself.

Step 3 — Assign it to Image Recognition

Open Settings → Agent → AI Formatting, find the Image recognition row, and set:

• Provider → Zhipu GLM
• Model → glm-4.6v-flash

Step 4 — Test

Start a DeepSeek session in Scholardo, paste in an image, and confirm it is described correctly.

Available models

Model	Notes
glm-4.6v-flash	Free, recommended for most use cases
glm-4.6v-flashx	9B, faster and more consistent
glm-4.6v	106B, highest quality

How it works

You paste an image into a DeepSeek session
        ↓
Local vision proxy (Python, launched with DeepSeek)
  intercepts image requests
        ↓
Calls Zhipu GLM-4.6V → image described as text
        ↓
Text description forwarded to DeepSeek
        ↓
DeepSeek replies based on the description

FAQ

Is glm-4.6v-flash good enough?
Yes for "what's in this image." For complex charts or formulas, upgrade to glm-4.6v-flashx or glm-4.6v.

Getting errors or images ignored?
Make sure you selected a vision model (glm-4.6v* series). Plain text models don't accept images. Also double-check the Base URL has no trailing /chat/completions.

Will my images be sent to the cloud?
Yes — images are sent to Zhipu's servers for transcription. Avoid sending sensitive content if that's a concern. For fully local processing, you can configure an Ollama vision model instead (higher local memory usage).