MineBench local lab

Test models locally
Test out changing the system prompt, generate a custom build, then paste the JSON to render it.
Grid size
Grid size

Sets the build volume. Larger grids give the model more room for detail and variation, which broadens the output range and helps keep the benchmark from getting saturated.

Block palette
Block palette

Sets the selection of blocks available to the model. Simple keeps the block set tight. Advanced unlocks a wider palette for richer, more varied builds.

System prompt
Here's the default system prompt the official benchmark uses. Feel free to play around with it.
User prompt
The actual object you want the model to build.
## Build Request
Here is what you need to build:
<build_request>
A warm wooden cabin beside a pond, with a stone chimney, a small dock, and a few trees.
</build_request>
Test Local Models or Try Directly On a Website (e.g. ChatGPT or Claude)
Press the Copy Both button to copy the system prompt with your user prompt.
Note: If you're generating the build through a site like chatgpt.com directly, add one final line asking for a downloadable JSON file or artifact attachment instead of raw JSON text. Otherwise the model will output just raw text and hit it's output limit.
Return only the final voxel object as a JSON file/artifact attachment.
Render JSON
Paste model output and render with Cmd/Ctrl+Enter.
Preview
Export GIF
No build yet