imageModel config for vision routing in OpenClaw 2026.4.26

[ruby egret]

Hi Krill! Running OpenClaw 2026.4.26 on a VPS with DeepSeek as primary model. We want to route image inputs automatically to Gemini (google/gemini-2.0-flash-001) while keeping DeepSeek for all text. We've added an imageModel key to the config but are unsure of the exact correct path. What is the correct openclaw.json config to set a separate model for image/vision inputs?

[mild parrot]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[snow meadow]

Yep — the path is:

{
  agents: {
    defaults: {
      model: {
        primary: "deepseek/<your-text-model>"
      },
      imageModel: {
        primary: "google/gemini-2.0-flash-001"
      }
    }
  }
}

Short form is also accepted:

{
  agents: {
    defaults: {
      model: "deepseek/<your-text-model>",
      imageModel: "google/gemini-2.0-flash-001"
    }
  }
}

agents.defaults.imageModel is the vision fallback used when the active/default reply model can’t accept images. So with DeepSeek as the text model, image attachments should route through Gemini for understanding.

Make sure the Google/Gemini auth is configured too (GEMINI_API_KEY or GOOGLE_API_KEY, depending on your setup), then restart/reload the gateway. You can also set it with:

openclaw models set-image google/gemini-2.0-flash-001
openclaw models status

Docs: https://docs.openclaw.ai/gateway/config-agents and https://docs.openclaw.ai/concepts/models

Tiny naming trap: this is for image/vision understanding. imageGenerationModel is separate and only for generating/editing images.