diff --git a/CorridorKeyModule/backend.py b/CorridorKeyModule/backend.py
index e3251318..fa803ab8 100644
--- a/CorridorKeyModule/backend.py
+++ b/CorridorKeyModule/backend.py
@@ -372,12 +372,16 @@ def create_engine(
     backend: str | None = None,
     device: str | None = None,
     img_size: int = DEFAULT_IMG_SIZE,
+    model_precision: torch.dtype = torch.float16,
     tile_size: int | None = DEFAULT_MLX_TILE_SIZE,
     overlap: int = DEFAULT_MLX_TILE_OVERLAP,
 ):
     """Factory: returns an engine with process_frame() matching the Torch contract.
 
     Args:
+        model_precision: Torch only. Model weight dtype. Defaults to
+            torch.float16 for lower VRAM. Override with torch.float32 if
+            numerical parity with eager CPU inference is required.
         tile_size: MLX only — tile size for tiled inference (default 512).
             Set to None to disable tiling and use full-frame inference.
         overlap: MLX only — overlap pixels between tiles (default 64).
@@ -398,5 +402,5 @@ def create_engine(
 
         logger.info("Torch engine loaded: %s (device=%s)", ckpt.name, device)
         return CorridorKeyEngine(
-            checkpoint_path=str(ckpt), device=device or "cpu", img_size=img_size, model_precision=torch.float16
+            checkpoint_path=str(ckpt), device=device or "cpu", img_size=img_size, model_precision=model_precision
         )
diff --git a/backend/service.py b/backend/service.py
index b7b1eb9f..81a37eb6 100644
--- a/backend/service.py
+++ b/backend/service.py
@@ -284,25 +284,25 @@ def _ensure_model(self, needed: _ActiveModel) -> None:
         self._active_model = needed
 
     def _get_engine(self):
-        """Lazy-load the CorridorKey inference engine."""
+        """Lazy-load the CorridorKey inference engine via the backend factory."""
         self._ensure_model(_ActiveModel.INFERENCE)
 
         if self._engine is not None:
             return self._engine
 
         try:
-            from CorridorKeyModule.backend import TORCH_EXT, _discover_checkpoint
-            from CorridorKeyModule.inference_engine import CorridorKeyEngine
+            import torch
+
+            from CorridorKeyModule.backend import create_engine
         except ImportError as exc:
             raise RuntimeError("CorridorKeyModule is not installed. Run: uv sync") from exc
 
-        ckpt_path = _discover_checkpoint(TORCH_EXT)
-        logger.info(f"Loading checkpoint: {os.path.basename(ckpt_path)}")
         t0 = time.monotonic()
-        self._engine = CorridorKeyEngine(
-            checkpoint_path=ckpt_path,
+        self._engine = create_engine(
+            backend="torch",
             device=self._device,
             img_size=2048,
+            model_precision=torch.float32,
         )
         logger.info(f"Engine loaded in {time.monotonic() - t0:.1f}s")
         return self._engine
diff --git a/tests/test_backend.py b/tests/test_backend.py
index 64e6b055..01ebc864 100644
--- a/tests/test_backend.py
+++ b/tests/test_backend.py
@@ -1,12 +1,14 @@
 """Unit tests for CorridorKeyModule.backend — no GPU/MLX required."""
 
 import errno
+import inspect
 import logging
 import os
 from unittest import mock
 
 import numpy as np
 import pytest
+import torch
 
 from CorridorKeyModule.backend import (
     BACKEND_ENV_VAR,
@@ -344,3 +346,65 @@ def test_value_ranges(self, mlx_raw_output):
         # (same behavior as Torch engine — linear_to_srgb doesn't clamp)
         for key in ("comp", "processed"):
             assert result[key].min() >= 0.0, f"{key} has negative values"
+
+
+# --- create_engine model_precision parameter ---
+
+
+class TestCreateEngineModelPrecision:
+    """create_engine exposes and forwards a model_precision parameter for Torch."""
+
+    def test_signature_has_model_precision(self):
+        """create_engine's signature includes a model_precision parameter."""
+        from CorridorKeyModule.backend import create_engine
+
+        params = inspect.signature(create_engine).parameters
+        assert "model_precision" in params
+
+    def test_default_model_precision_is_float16(self):
+        """Default model_precision is torch.float16 to keep CLI and service aligned."""
+        from CorridorKeyModule.backend import create_engine
+
+        params = inspect.signature(create_engine).parameters
+        assert params["model_precision"].default is torch.float16
+
+    def test_model_precision_forwarded_to_torch_engine(self, tmp_path):
+        """create_engine forwards the requested model_precision to CorridorKeyEngine."""
+        from CorridorKeyModule.backend import create_engine
+
+        ckpt = tmp_path / "model.safetensors"
+        ckpt.touch()
+        with mock.patch("CorridorKeyModule.backend.CHECKPOINT_DIR", str(tmp_path)):
+            with mock.patch("CorridorKeyModule.inference_engine.CorridorKeyEngine") as mock_engine_cls:
+                create_engine(backend="torch", device="cpu", model_precision=torch.float32)
+                mock_engine_cls.assert_called_once()
+                kwargs = mock_engine_cls.call_args.kwargs
+                assert kwargs["model_precision"] is torch.float32
+
+
+# --- Service routes engine construction through the factory ---
+
+
+class TestServiceEngineRouting:
+    """CorridorKeyService._get_engine delegates engine construction to create_engine."""
+
+    def test_get_engine_calls_create_engine_with_torch_and_fp32(self):
+        """_get_engine invokes the factory with explicit backend="torch" and FP32 precision.
+
+        FP32 is passed explicitly to preserve the service's existing quality-first
+        behavior (FP32 weights plus mixed_precision autocast for speed on safe ops).
+        The factory's default FP16 matches the CLI's long-standing VRAM tradeoff.
+        """
+        from backend.service import CorridorKeyService
+
+        service = CorridorKeyService()
+        with mock.patch("CorridorKeyModule.backend.create_engine") as mock_factory:
+            mock_factory.return_value = object()
+            service._get_engine()
+
+            mock_factory.assert_called_once()
+            kwargs = mock_factory.call_args.kwargs
+            assert kwargs["backend"] == "torch"
+            assert kwargs["device"] == "cpu"
+            assert kwargs["img_size"] == 2048
+            assert kwargs["model_precision"] is torch.float32