Setting up local Gemma4 26B A4B

[broken basin]

what should i change in the openclaw.json to make it worl perfectly?
"models": {
"mode": "merge",
"providers": {
"local-gemma": {
"baseUrl": "http://127.0.0.1:8082/v1",
"apiKey": "not-needed",
"api": "openai-completions",
"models": [
{
"id": "Gemma4-26B-A4B",
"name": "Google Gemma 4 26B A4B",
"reasoning": false,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 262144,
"maxTokens": 8192
}
]
},

[gusty dew]

Not “perfectly” 😅, but for the best chance with a local Gemma OpenAI-compatible backend, I’d change/add these:

{
  agents: {
    defaults: {
      model: { primary: "local-gemma/Gemma4-26B-A4B" },
      models: {
        "local-gemma/Gemma4-26B-A4B": { alias: "Gemma" },
      },
    },
  },

  models: {
    mode: "merge",
    providers: {

      "local-gemma": {
        baseUrl: "http://127.0.0.1:8082/v1",
        apiKey: "gemma-local",
        api: "openai-completions",
        models: [
          {
            id: "Gemma4-26B-A4B",
            name: "Google Gemma 4 26B A4B",
            reasoning: false,
            input: ["text"],
            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
            contextWindow: 262144,
            maxTokens: 8192,
            compat: {
              requiresStringContent: true,

              supportsTools: false
            }
          }
        ]
      }
    }
  }
}

Most important bits:

• agents.defaults.model.primary so OpenClaw actually uses it
• compat.requiresStringContent: true for stricter local /v1 backends
• compat.supportsTools: false because Gemma/local OpenAI-compatible servers often break on full agent tool schemas

Also, if you already have an agents.defaults.models allowlist elsewhere, add this model there too, or it may be rejected as “not allowed”.

Please paste these too:

openclaw --version
openclaw models status

And if your server is specifically vLLM, SGLang, LM Studio, or llama.cpp, say which one, because the best config differs a bit.

Docs:

• https://docs.openclaw.ai/gateway/local-models
• https://docs.openclaw.ai/help/troubleshooting
• https://docs.openclaw.ai/concepts/models

[broken basin]

it should support not only text but even other formats, like image and audio

[gusty dew]

Yep, but there’s an important catch:

• input in models.providers...models[] only supports "text" and "image"
• "audio" is not valid there
• Audio support is configured separately under tools.media.audio

So for your local Gemma block, the multimodal version would look more like:

{
  agents: {
    defaults: {
      model: { primary: "local-gemma/Gemma4-26B-A4B" },
    },
  },

  models: {
    mode: "merge",
    providers: {
      "local-gemma": {
        baseUrl: "http://127.0.0.1:8082/v1",
        apiKey: "gemma-local",
        api: "openai-completions",
        models: [
          {
            id: "Gemma4-26B-A4B",
            name: "Google Gemma 4 26B A4B",
            reasoning: false,
            input: ["text", "image"],
            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
            contextWindow: 262144,

            maxTokens: 8192,
            compat: {
              requiresStringContent: true,
              supportsTools: false
            }
          }
        ]
      }
    }
  },

  tools: {
    media: {
      image: {
        enabled: true

      },
      audio: {
        enabled: true
      }
    }
  }
}

Two important notes:

1. Image

   • input: ["text", "image"] is correct only if your Gemma server on :8082 actually exposes vision
   • if the backend is text-only, config alone won’t make images work

2. Audio

• audio is not handled by input
  • use tools.media.audio
  • if you want reliable audio, you usually add a transcription model or CLI fallback, for example OpenAI/Groq/Deepgram/Whisper

So the short answer is:

• change input to ["text", "image"] for image support
• add tools.media.audio.enabled: true for audio processing
• do not add "audio" to the model input array

If you want, paste your full openclaw.json plus:

openclaw --version
openclaw models status