diff --git a/autoarray/operators/convolver.py b/autoarray/operators/convolver.py
index ef793f54..702d1f16 100644
--- a/autoarray/operators/convolver.py
+++ b/autoarray/operators/convolver.py
@@ -699,7 +699,10 @@ def from_gaussian(
         Parameters
         ----------
         shape_native
-            The 2D shape of the mask the array is paired with.
+            The 2D shape of the mask the array is paired with. The kernel is always built at this
+            size, including when ``PYAUTO_SMALL_DATASETS=1`` is set — a kernel's shape is intrinsic
+            to the convolution operator, not a dataset size, so the fast-mode dataset cap does not
+            apply to it.
         pixel_scales
             The (y,x) arcsecond-to-pixel units conversion factor of every pixel. If this is input as a `float`,
             it is converted to a (float, float).
@@ -718,7 +721,13 @@ def from_gaussian(
             ``pixel_scales`` input should be the fine resolution (image pixel scale divided by this size).
         """
 
-        grid = Grid2D.uniform(shape_native=shape_native, pixel_scales=pixel_scales)
+        grid = Grid2D.uniform(
+            shape_native=shape_native,
+            pixel_scales=pixel_scales,
+            # The kernel is wrapped in an `Array2D` at `shape_native` below, so letting the
+            # `PYAUTO_SMALL_DATASETS` cap shrink this grid would leave the two inconsistent.
+            respect_small_datasets=False,
+        )
         grid_shifted = np.subtract(grid.array, centre)
         grid_radius = np.sqrt(np.sum(grid_shifted**2.0, 1))
         theta_coordinate_to_profile = np.arctan2(
diff --git a/test_autoarray/operators/test_convolver.py b/test_autoarray/operators/test_convolver.py
index 70f26466..43adfab6 100644
--- a/test_autoarray/operators/test_convolver.py
+++ b/test_autoarray/operators/test_convolver.py
@@ -53,6 +53,30 @@ def test__from_gaussian__normalized__kernel_values_match_expected():
     )
 
 
+def test__from_gaussian__small_datasets__kernel_keeps_full_requested_shape(monkeypatch):
+    """
+    The `PYAUTO_SMALL_DATASETS` fast-mode cap shrinks datasets (and the grids paired with
+    them) to 16x16, but a convolution kernel's shape is intrinsic to the operator. If the
+    cap reached the grid the Gaussian is evaluated on, its 256 values would be wrapped in an
+    `Array2D` at the uncapped `shape_native` of 961, raising an `ArrayException`.
+    """
+    monkeypatch.setenv("PYAUTO_SMALL_DATASETS", "1")
+
+    convolver = aa.Convolver.from_gaussian(
+        shape_native=(31, 31),
+        pixel_scales=0.1,
+        sigma=0.1,
+    )
+
+    assert convolver.kernel.shape_native == (31, 31)
+
+    image = aa.Array2D.ones(shape_native=(16, 16), pixel_scales=0.1)
+
+    blurred_image = convolver.convolved_image_from(image=image, blurring_image=None)
+
+    assert blurred_image.shape_native == (16, 16)
+
+
 def test__normalize__ones_kernel__each_element_equals_one_ninth():
 
     kernel_data = aa.Array2D.ones(shape_native=(3, 3), pixel_scales=1.0)