Using Docker Sandboxes (sbx) with local models via Docker Model Runner

sbx-local-models.md

Using Docker Sandboxes (sbx) with Local Models via Docker Model Runner

This works today, even before first-class support lands. Example uses opencode with Docker Model Runner (DMR).

Steps

1. Pull a model

docker model pull qwen3.6

2. (Optional) Allow network access to DMR

host.docker.internal is allowed by default, so this is usually not needed. If you hit a network policy issue, run:

sbx policy allow network localhost:12434

3. Configure your agent to talk to DMR

For opencode, drop this opencode.json into the root of your project:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "qwen3.6",
  "provider": {
    "dmr": {
      "models": {
        "qwen3.6": {
          "name": "qwen3.6"
        }
      },
      "name": "Docker Model Runner",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://host.docker.internal:12434/v1"
      }
    }
  }
}

4. Run the sandbox

sbx run opencode

That's it.

Note on context window

You don't need to override the context window manually. DMR no longer caps context at 4k by default — models advertise their real context length. For example:

docker model inspect -r qwen3.6
# "qwen35moe.context_length": "262144"

---

Credit: thanks to Ignasi Lopez Luna and Chris Crone for the recipe.