diff --git a/test/stdlib/test_spans.py b/test/stdlib/test_spans.py
index 06c9f8a2f..82e63c547 100644
--- a/test/stdlib/test_spans.py
+++ b/test/stdlib/test_spans.py
@@ -8,7 +8,12 @@
 from mellea.stdlib.session import MelleaSession, start_session
 
 # Module-level markers for all tests using Granite 4 hybrid micro (3B model)
-pytestmark = [pytest.mark.huggingface, pytest.mark.requires_gpu, pytest.mark.llm]
+pytestmark = [
+    pytest.mark.huggingface,
+    pytest.mark.requires_gpu,
+    pytest.mark.requires_heavy_ram,
+    pytest.mark.llm,
+]
 
 
 # We edit the context type in the async tests below. Don't change the scope here.
diff --git a/test/telemetry/test_metrics_backend.py b/test/telemetry/test_metrics_backend.py
index 9bba4503a..5b2702bd1 100644
--- a/test/telemetry/test_metrics_backend.py
+++ b/test/telemetry/test_metrics_backend.py
@@ -51,6 +51,33 @@ def enable_metrics(monkeypatch):
     importlib.reload(mellea.telemetry.metrics)
 
 
+@pytest.fixture(scope="module")
+def hf_metrics_backend(gh_run):
+    """Shared HuggingFace backend for telemetry metrics tests.
+
+    Uses module scope to load the model once and reuse it across all tests,
+    preventing memory exhaustion from loading multiple model instances.
+    """
+    if gh_run:
+        pytest.skip("Skipping HuggingFace backend creation in CI")
+
+    from mellea.backends.cache import SimpleLRUCache
+    from mellea.backends.huggingface import LocalHFBackend
+
+    backend = LocalHFBackend(
+        model_id=IBM_GRANITE_4_HYBRID_MICRO.hf_model_name,  # type: ignore
+        cache=SimpleLRUCache(5),
+    )
+
+    yield backend
+
+    # Cleanup
+    import gc
+
+    del backend
+    gc.collect()
+
+
 def get_metric_value(metrics_data, metric_name, attributes=None):
     """Helper to extract metric value from metrics data.
 
@@ -297,15 +324,12 @@ async def test_litellm_token_metrics_integration(
 @pytest.mark.asyncio
 @pytest.mark.llm
 @pytest.mark.huggingface
+@pytest.mark.requires_heavy_ram
 @pytest.mark.parametrize("stream", [False, True], ids=["non-streaming", "streaming"])
 async def test_huggingface_token_metrics_integration(
-    enable_metrics, metric_reader, stream, gh_run
+    enable_metrics, metric_reader, stream, hf_metrics_backend
 ):
     """Test that HuggingFace backend records token metrics correctly."""
-    if gh_run:
-        pytest.skip("Skipping in CI - requires model download")
-
-    from mellea.backends.huggingface import LocalHFBackend
     from mellea.backends.model_options import ModelOption
     from mellea.telemetry import metrics as metrics_module
 
@@ -315,17 +339,11 @@ async def test_huggingface_token_metrics_integration(
     metrics_module._input_token_counter = None
     metrics_module._output_token_counter = None
 
-    from mellea.backends.cache import SimpleLRUCache
-
-    backend = LocalHFBackend(
-        model_id=IBM_GRANITE_4_HYBRID_MICRO.hf_model_name,  # type: ignore
-        cache=SimpleLRUCache(5),
-    )
     ctx = SimpleContext()
     ctx = ctx.add(Message(role="user", content="Say 'hello' and nothing else"))
 
     model_options = {ModelOption.STREAM: True} if stream else {}
-    mot, _ = await backend.generate_from_context(
+    mot, _ = await hf_metrics_backend.generate_from_context(
         Message(role="assistant", content=""), ctx, model_options=model_options
     )