diff --git a/src/markovian_rsa_mlx/__init__.py b/src/markovian_rsa_mlx/__init__.py
index d642c98..e5ae407 100644
--- a/src/markovian_rsa_mlx/__init__.py
+++ b/src/markovian_rsa_mlx/__init__.py
@@ -1,2 +1,7 @@
 """Markovian RSA test-time compute methodology on MLX."""
 __version__ = "0.1.0"
+
+from markovian_rsa_mlx.config import RSAConfig
+from markovian_rsa_mlx.loader import load_zaya_model
+
+__all__ = ["__version__", "RSAConfig", "load_zaya_model"]
diff --git a/src/markovian_rsa_mlx/loader.py b/src/markovian_rsa_mlx/loader.py
new file mode 100644
index 0000000..2b3eebd
--- /dev/null
+++ b/src/markovian_rsa_mlx/loader.py
@@ -0,0 +1,22 @@
+"""Model + tokenizer loading via mlx-lm. Local paths and HF repo ids supported."""
+from __future__ import annotations
+from pathlib import Path
+from typing import Any
+
+
+def resolve_model_path(model_id: str) -> str:
+    """Pass-through resolver. mlx-lm handles HF download for repo ids."""
+    p = Path(model_id)
+    if p.exists():
+        return str(p)
+    return model_id
+
+
+def load_zaya_model(model_id: str, *, trust_remote_code: bool = True) -> tuple[Any, Any]:
+    """Load a ZAYA-supporting MLX checkpoint via mlx-lm.
+
+    Requires the kyr0/mlx-lm fork (feat/zaya-support) which adds the `zaya`
+    architecture support. Returns (model, tokenizer).
+    """
+    from mlx_lm import load
+    return load(resolve_model_path(model_id), tokenizer_config={"trust_remote_code": trust_remote_code})
diff --git a/tests/test_loader.py b/tests/test_loader.py
new file mode 100644
index 0000000..cb4f8df
--- /dev/null
+++ b/tests/test_loader.py
@@ -0,0 +1,25 @@
+import pytest
+from markovian_rsa_mlx.loader import resolve_model_path
+
+
+def test_resolve_local_path_passthrough(tmp_path):
+    # local dir is returned as-is
+    d = tmp_path / "weights"
+    d.mkdir()
+    assert resolve_model_path(str(d)) == str(d)
+
+
+def test_resolve_hf_id_format_unchanged():
+    # HF repo id is returned as-is (load_model handles download)
+    assert resolve_model_path("kyr0/zaya1-base-8b-MLX") == "kyr0/zaya1-base-8b-MLX"
+
+
+@pytest.mark.integration
+def test_load_real_zaya_q4_smoke():
+    """Hits the actual Q4 weights — slow, requires HF cache."""
+    from markovian_rsa_mlx.loader import load_zaya_model
+    model, tokenizer = load_zaya_model("kyr0/zaya1-base-8b-MLX")
+    assert model is not None
+    assert tokenizer is not None
+    assert hasattr(tokenizer, "encode")
+    assert hasattr(tokenizer, "decode")