diff --git a/docker-compose.yml b/docker-compose.yml
index f8b6d6498..794e6ecbc 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -9,6 +9,7 @@ volumes:
   minio_data:
     driver: local
   node_modules:
+  ml_backend_model_cache:
 
 services:
   django: &django
@@ -187,9 +188,11 @@ services:
 
   ml_backend:
     build:
-      context: ./processing_services/minimal
+      context: ./processing_services/example
     volumes:
-      - ./processing_services/minimal/:/app
+      - ./processing_services/example/:/app
+      # Hugging Face and torch downloads survive container rebuilds; the first start downloads the models.
+      - ml_backend_model_cache:/root/.cache
     ports:
       - "2005:2000"
     networks:
diff --git a/processing_services/README.md b/processing_services/README.md
index 48bb254e7..cf61c8315 100644
--- a/processing_services/README.md
+++ b/processing_services/README.md
@@ -11,8 +11,8 @@ In this directory, we define locally-run processing services as FastAPI apps. A
 - `/process`: receives source images and existing detections via a `PipelineRequest` and returns a `PipelineResponse` containing detections
 
 `processing_services` contains 2 apps:
-- `example`: demos how to add custom pipelines/algorithms.
-- `minimal`: a simple ML backend for basic testing of the processing service API. This minimal app also runs within the main Antenna docker compose stack.
+- `example`: demos how to add custom pipelines/algorithms with real models. It runs as `ml_backend` in the main Antenna docker compose stack for local development, so jobs there produce real detections and classifications. The first job downloads the models (about 3.6 GB for the zero-shot pipelines; about a minute on a fast connection) into the `ml_backend_model_cache` volume, which keeps them across rebuilds; `docker volume rm` it to download again.
+- `minimal`: a simple ML backend with no models and fixed outputs, for testing the processing service API. The CI stack (`docker-compose.ci.yml`) uses it, and it can be started next to the main stack with `docker compose -f processing_services/docker-compose.yml up -d ml_backend_minimal`.
 
 If your goal is to run an ML backend locally, simply copy the `example` app and follow the steps below.