diff --git a/pyproject.toml b/pyproject.toml
index 47e1a8b6c..c04cf0eec 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -5,7 +5,7 @@ requires = ["setuptools >= 61.0", "wheel"]
 [project]
 name = "trimesh"
 requires-python = ">=3.7"
-version = "4.0.5"
+version = "4.0.6"
 authors = [{name = "Michael Dawson-Haggerty", email = "mikedh@kerfed.com"}]
 license = {file = "LICENSE.md"}
 description = "Import, export, process, analyze and view triangular meshes."
diff --git a/tests/test_inertia.py b/tests/test_inertia.py
index 8b96b7754..86797200a 100644
--- a/tests/test_inertia.py
+++ b/tests/test_inertia.py
@@ -43,7 +43,10 @@ def test_inertia(self):
         assert g.np.abs(g.np.dot(t1, t1.T) - g.np.eye(4)).max() < 1e-10
 
         c = g.trimesh.primitives.Cylinder(
-            height=10, radius=1, sections=720, transform=t0  # number of slices
+            height=10,
+            radius=1,
+            sections=720,
+            transform=t0,  # number of slices
         )
         c0m = c.moment_inertia.copy()
         c0 = g.trimesh.inertia.cylinder_inertia(
@@ -343,9 +346,7 @@ def parallel_axis_theorem(inertia, mass, a1, a2, a3):
         # CHECK FRAME 0
         # analytical calculations of inertia tensor by hand
         inertia0 = (
-            0.083333333333
-            * mass
-            * g.np.diag([h**2 + b**2, h**2 + d**2, b**2 + d**2])
+            0.083333333333 * mass * g.np.diag([h**2 + b**2, h**2 + d**2, b**2 + d**2])
         )
         a1 = -0.5 * d
         a2 = -0.5 * b
@@ -358,9 +359,7 @@ def parallel_axis_theorem(inertia, mass, a1, a2, a3):
         # CHECK FRAME 1
         # analytical calculations of inertia tensor by hand
         inertia1 = (
-            0.083333333333
-            * mass
-            * g.np.diag([h**2 + d**2, h**2 + b**2, b**2 + d**2])
+            0.083333333333 * mass * g.np.diag([h**2 + d**2, h**2 + b**2, b**2 + d**2])
         )
         a1 = -0.5 * b
         a2 = 0.5 * d
@@ -376,9 +375,7 @@ def parallel_axis_theorem(inertia, mass, a1, a2, a3):
         # CHECK FRAME 2
         # analytical calculations of inertia tensor by hand
         inertia2 = (
-            0.083333333333
-            * mass
-            * g.np.diag([h**2 + b**2, b**2 + d**2, h**2 + d**2])
+            0.083333333333 * mass * g.np.diag([h**2 + b**2, b**2 + d**2, h**2 + d**2])
         )
         a1 = -0.5 * d
         a2 = 0.5 * h
diff --git a/trimesh/grouping.py b/trimesh/grouping.py
index 40018937d..0c8d6fc7c 100644
--- a/trimesh/grouping.py
+++ b/trimesh/grouping.py
@@ -189,6 +189,29 @@ def hashable_rows(data, digits=None):
     if len(as_int.shape) == 1:
         return as_int
 
+
+    # if array is 2D and smallish, we can try bitbanging
+    # this is significantly faster than the custom dtype
+    if len(as_int.shape) == 2 and as_int.shape[1] <= 4:
+        # time for some righteous bitbanging
+        # can we pack the whole row into a single 64 bit integer
+        precision = int(np.floor(64 / as_int.shape[1]))
+
+        d_min, d_max = as_int.min(), as_int.max()
+
+        # if the max value is less than precision we can do this
+        if abs(d_max - d_min) < 2 ** (precision - 1):
+            # the resulting package
+            hashable = np.zeros(len(as_int), dtype=np.uint64)
+            # offset to zero
+            bitbang = (as_int - d_min).astype(np.uint64).T
+            # loop through each column and bitwise xor to combine
+            # make sure as_int is int64 otherwise bit offset won't work
+            for offset, column in enumerate(bitbang)
+                # will modify hashable in place
+                np.bitwise_xor(hashable, column << (offset * precision), out=hashable)
+            return hashable
+
     # reshape array into magical data type that is weird but hashable
     dtype = np.dtype((np.void, as_int.dtype.itemsize * as_int.shape[1]))