The smallest example, and the one the setup guide runs. Its application lists the catalogue through the API, submits one request to every catalogued model, polls each to completion, prints a line per model, and exits 0 only if every model answered.
Run
marigold package create marigold-examples/platform-model-test -o /tmp
marigold package install /tmp/platform-model-test-0.2.0.tar.gz
marigold cache populate platform-model-test
marigold application start platform-model-test
marigold application logs platform-model-test
Models
qwen/qwen3-0.6b– instructhuggingfacetb/smollm2-135m– instructsentence-transformers/all-minilm-l6-v2– text-embeddingopenai/clip-vit-base-patch32– image-embeddinghuggingfacetb/smolvlm-256m-instruct– img2txt
Result
marigold application status platform-model-test
The exit code is 0 if every model answered. The catalogue is host-wide, so the application tests every cached model, including models cached for other packages. The first requests are slow while each model loads.
The application contract
main.py is the reference for writing an application. It reads
MARIGOLD_API_BASE and MARIGOLD_APPLICATION_ID from its environment,
sets application_id and a fresh nonce on every request, and polls
the Location each submission returns until the job completes.
Stop
marigold application stop platform-model-test