diff --git a/brainscore_language/models/phi2/__init__.py b/brainscore_language/models/phi2/__init__.py
new file mode 100644
index 00000000..ede23158
--- /dev/null
+++ b/brainscore_language/models/phi2/__init__.py
@@ -0,0 +1,12 @@
+from brainscore_language import model_registry
+from brainscore_language import ArtificialSubject
+from brainscore_language.model_helpers.huggingface import HuggingfaceSubject
+
+# Phi-2: 32 transformer layers, hidden size 2560.
+# Layer 31 (last) chosen as default mapping pending benchmark-driven selection.
+model_registry['phi-2'] = lambda: HuggingfaceSubject(
+    model_id='microsoft/phi-2',
+    region_layer_mapping={
+        ArtificialSubject.RecordingTarget.language_system: 'model.layers.31'
+    },
+)
diff --git a/brainscore_language/models/phi2/metadata.yml b/brainscore_language/models/phi2/metadata.yml
new file mode 100644
index 00000000..589fe19b
--- /dev/null
+++ b/brainscore_language/models/phi2/metadata.yml
@@ -0,0 +1,15 @@
+models:
+  phi2:
+    architecture: DCNN
+    model_family: phi2
+    total_parameter_count: 1234567
+    trainable_parameter_count: 1234567
+    total_layers: 55
+    trainable_layers: 40
+    model_size_mb: 1202
+    training_dataset: null
+    task_specialization: null
+    brainscore_link: https://github.com/brain-score/language/tree/master/brainscore_language/models/phi2
+    huggingface_link: null
+    extra_notes: Temporary hardcoded metadata - will be replaced with actual generation
+    runnable: true
diff --git a/brainscore_language/models/phi2/test.py b/brainscore_language/models/phi2/test.py
new file mode 100644
index 00000000..bad59f9d
--- /dev/null
+++ b/brainscore_language/models/phi2/test.py
@@ -0,0 +1,34 @@
+import numpy as np
+import pytest
+
+from brainscore_language import load_model
+from brainscore_language.artificial_subject import ArtificialSubject
+
+
+@pytest.mark.memory_intense
+def test_load_model():
+    """Model can be loaded from the registry without errors."""
+    model = load_model('phi-2')
+    assert model is not None
+
+
+@pytest.mark.memory_intense
+def test_identifier():
+    model = load_model('phi-2')
+    assert model.identifier() == 'microsoft/phi-2'
+
+
+@pytest.mark.memory_intense
+def test_neural():
+    """Model produces neural representations with the expected shape."""
+    model = load_model('phi-2')
+    text = ['the quick brown fox', 'jumps over', 'the lazy dog']
+    model.start_neural_recording(
+        recording_target=ArtificialSubject.RecordingTarget.language_system,
+        recording_type=ArtificialSubject.RecordingType.fMRI,
+    )
+    representations = model.digest_text(text)['neural']
+    assert len(representations['presentation']) == 3
+    np.testing.assert_array_equal(representations['stimulus'], text)
+    assert len(representations['neuroid']) == 2560
+