Discover Axelera models

Start with precompiled Axelera models that run out of the box—learn how to pick the right variant for your device.

Estimated read time: 3 minutes

Where the model comes from: Axelera Model Zoo

In the First inference guide, we specified the model by:

• model_name="yolov8n_coco--640x640_quant_axelera_metis_1"

• zoo_url="degirum/axelera"

• inference_host_address="@local"

• model_properties={"device_type": ["AXELERA/METIS"]}

That zoo_url points to the Axelera Model Zoo, a curated collection of models precompiled for Axelera Metis hardware. We highlight this zoo because it lets you run examples immediately, without any conversion steps.

Any model compiled for Axelera can be used with the same pattern. The advanced guide shows how to bring your own model (BYOM).

Why we start with the Axelera Model Zoo

• Instant results: models are precompiled and follow predictable I/O, so examples “just work”

• Consistent setup: you focus on model_name, zoo_url, and device_type—everything else stays the same

• Easy to swap: change only the model_name to try a different model

Example models

# ImageNet classification model
model_spec = ModelSpec(
    model_name="yolov8n_imagenet--224x224_quant_axelera_metis_1",
    zoo_url="degirum/axelera",
    inference_host_address="@local",
    model_properties={"device_type": ["AXELERA/METIS"]},
)

# COCO object detection model
model_spec = ModelSpec(
    model_name="yolov8n_coco--640x640_quant_axelera_metis_1",
    zoo_url="degirum/axelera",
    inference_host_address="@local",
    model_properties={"device_type": ["AXELERA/METIS"]},
)

# Face detection model
model_spec = ModelSpec(
    model_name="yolov8n_relu6_face--640x640_quant_axelera_metis_1",
    zoo_url="degirum/axelera",
    inference_host_address="@local",
    model_properties={"device_type": ["AXELERA/METIS"]},
)

Bring your own Axelera model (advanced)

When you’re ready, the advanced guide shows how to:

• Package your Axelera-compiled artifacts with a PySDK model.json that defines inputs, outputs, preprocessing, and postprocessing.

• Point PySDK to your private or local model zoo.

• Rescan the zoo as you add models.

• Use the same ModelSpec pattern—only the zoo_url and model_name change.

Bottom line: Start fast with the Axelera Model Zoo, then graduate to bring in your own Axelera models, all using the same stable PySDK workflow.