diff --git a/pyro-predictor/pyro_predictor/vision.py b/pyro-predictor/pyro_predictor/vision.py
index 33166cfb..0e30bcb4 100644
--- a/pyro-predictor/pyro_predictor/vision.py
+++ b/pyro-predictor/pyro_predictor/vision.py
@@ -6,6 +6,7 @@
 import logging
 import pathlib
 import platform
+import shutil
 import tarfile
 from typing import Tuple
 
@@ -21,6 +22,8 @@
 
 MODEL_REPO_ID = "pyronear/yolo11s_plucky-pelican_v7.1.0"
 MODEL_NAME = "ncnn_cpu.tar.gz"
+MODEL_SLUG = MODEL_REPO_ID.split("/", 1)[1]  # e.g. "yolo11s_plucky-pelican_v7.1.0"
+MODEL_CACHE_SUBDIR = "models"
 
 logging.basicConfig(format="%(asctime)s | %(levelname)s: %(message)s", level=logging.INFO, force=True)
 logger = logging.getLogger(__name__)
@@ -70,18 +73,37 @@ def __init__(
             else:
                 raise ValueError("Unsupported format: should be 'ncnn' or 'onnx'")
 
-            model_path = str(pathlib.Path(model_folder) / model)
+            # Namespace cached weights by model slug so a MODEL_REPO_ID bump lands in a
+            # fresh path and old weights can be purged.
+            cache_root = pathlib.Path(model_folder) / MODEL_CACHE_SUBDIR
+            model_cache = cache_root / MODEL_SLUG
+            model_path = str(model_cache / model)
 
             if not pathlib.Path(model_path).is_file():
+                # Drop previous slugs and legacy flat layout to reclaim disk on edge devices.
+                if cache_root.is_dir():
+                    for entry in cache_root.iterdir():
+                        if entry.name != MODEL_SLUG:
+                            shutil.rmtree(entry, ignore_errors=True)
+                            logger.info(f"Removed stale model cache: {entry}")
+                legacy_archive = pathlib.Path(model_folder) / model
+                legacy_extract = pathlib.Path(model_folder) / model.replace(".tar.gz", "")
+                if legacy_archive.is_file():
+                    legacy_archive.unlink()
+                    logger.info(f"Removed legacy model archive: {legacy_archive}")
+                if legacy_extract.is_dir():
+                    shutil.rmtree(legacy_extract, ignore_errors=True)
+                    logger.info(f"Removed legacy model extract dir: {legacy_extract}")
+
                 logger.info(f"Downloading model from {MODEL_REPO_ID}/{model} ...")
-                pathlib.Path(model_folder).mkdir(exist_ok=True, parents=True)
-                hf_hub_download(repo_id=MODEL_REPO_ID, filename=model, local_dir=model_folder)
+                model_cache.mkdir(exist_ok=True, parents=True)
+                hf_hub_download(repo_id=MODEL_REPO_ID, filename=model, local_dir=str(model_cache))
                 logger.info("Model downloaded!")
 
             # Extract archive
             if model_path.endswith(".tar.gz"):
                 base_name = pathlib.Path(model_path).name.replace(".tar.gz", "")
-                extract_path = str(pathlib.Path(model_folder) / base_name)
+                extract_path = str(model_cache / base_name)
                 if not pathlib.Path(extract_path).is_dir():
                     pathlib.Path(extract_path).mkdir(parents=True, exist_ok=True)
                     with tarfile.open(model_path, "r:gz") as tar:
diff --git a/tests/test_vision.py b/tests/test_vision.py
index 2b9fc02b..5ac10e76 100644
--- a/tests/test_vision.py
+++ b/tests/test_vision.py
@@ -6,6 +6,7 @@
 
 # Canonical import — Classifier lives in pyro_predictor
 from pyro_predictor import Classifier
+from pyro_predictor.vision import MODEL_CACHE_SUBDIR, MODEL_SLUG
 
 # pyroengine.vision shim must re-export the same class
 from pyroengine.vision import Classifier as ClassifierShim
@@ -29,7 +30,7 @@ def test_classifier(tmpdir_factory, mock_wildfire_image):
 
     # Test onnx model
     model = Classifier(model_folder=folder, format="onnx")
-    model_path = str(pathlib.Path(folder) / "onnx_cpu" / "best.onnx")
+    model_path = str(pathlib.Path(folder) / MODEL_CACHE_SUBDIR / MODEL_SLUG / "onnx_cpu" / "best.onnx")
     assert pathlib.Path(model_path).is_file()
 
     # Test occlusion mask
@@ -49,12 +50,36 @@ def sha256sum(path):
     return hashlib.sha256(pathlib.Path(path).read_bytes()).hexdigest()
 
 
+def test_stale_cache_is_purged(tmpdir_factory):
+    folder = pathlib.Path(tmpdir_factory.mktemp("engine_cache"))
+    models_dir = folder / MODEL_CACHE_SUBDIR
+
+    # Seed a previous-version slug alongside the current one.
+    stale_slug = models_dir / "yolo11s_stale-slug_v0.0.0"
+    stale_slug.mkdir(parents=True)
+    (stale_slug / "marker.txt").write_text("stale")
+
+    # Seed the pre-slug flat layout (onnx variant, matching the format used below).
+    legacy_archive = folder / "onnx_cpu.tar.gz"
+    legacy_archive.write_bytes(b"legacy")
+    legacy_extract = folder / "onnx_cpu"
+    legacy_extract.mkdir()
+    (legacy_extract / "marker.txt").write_text("legacy")
+
+    _ = Classifier(model_folder=str(folder), format="onnx")
+
+    assert not stale_slug.exists(), "stale slug should have been purged"
+    assert not legacy_archive.exists(), "legacy archive should have been purged"
+    assert not legacy_extract.exists(), "legacy extract dir should have been purged"
+    assert (models_dir / MODEL_SLUG / "onnx_cpu" / "best.onnx").is_file()
+
+
 def test_download(tmpdir_factory):
     folder = str(tmpdir_factory.mktemp("engine_cache"))
 
     # First download
     _ = Classifier(model_folder=folder, format="onnx")
-    model_path = str(pathlib.Path(folder) / "onnx_cpu" / "best.onnx")
+    model_path = str(pathlib.Path(folder) / MODEL_CACHE_SUBDIR / MODEL_SLUG / "onnx_cpu" / "best.onnx")
     assert pathlib.Path(model_path).is_file()
 
     hash1 = sha256sum(model_path)