diff --git a/coreai_torch/_aten_to_core.py b/coreai_torch/_aten_to_core.py
index 20a2f83..e589436 100644
--- a/coreai_torch/_aten_to_core.py
+++ b/coreai_torch/_aten_to_core.py
@@ -644,15 +644,16 @@ def replace_arange_start_step(
         else coreai.constant(1, dtype=start.type.element_type)
     )
 
-    # When operands are integer-typed, keep them as si32 so coreai.range_ can
-    # infer a static output shape, then cast the result to the requested dtype.
-    # For float operands the count isn't statically determinable anyway, so
-    # cast everything to target_type and let range_ return a dynamic shape.
+    # When ALL operands are integer-typed, keep them as si32 so coreai.range_
+    # can infer a static output shape, then cast the result to the requested
+    # dtype. If any operand is float (e.g. arange(0, 5, 0.5)), fall back to
+    # target_type — truncating a float step to si32 would corrupt the values.
     target_type = get_output_element_type_from_node(node)
     si32 = IntegerType.get_signed(32)
-    range_type = (
-        si32 if isinstance(start.type.element_type, IntegerType) else target_type
+    all_integer = all(
+        isinstance(v.type.element_type, IntegerType) for v in (start, end, step)
     )
+    range_type = si32 if all_integer else target_type
 
     def to_scalar(v: Value) -> Value:
         if v.type.rank > 0:
diff --git a/tests/ops/test_ops.py b/tests/ops/test_ops.py
index 6e85b7b..e244d28 100644
--- a/tests/ops/test_ops.py
+++ b/tests/ops/test_ops.py
@@ -347,7 +347,20 @@ def forward(self, x: Tensor) -> Tensor:
         x = torch.zeros(2, 8)
         await validate_numerical_output(model=model, x=x)
 
-    async def test_symint_end_with_float_start_step(self) -> None:
+    async def test_mixed_int_float_operands(self) -> None:
+        """Regression: arange with mixed int/float operands must not truncate.
+
+        torch.arange(0, 5, 0.5) has int start/end but float step. The lowering
+        must not cast step to si32 (which would truncate 0.5 → 0 and produce a
+        degenerate range); it must fall back to the float path instead.
+        """
+
+        class ArangeMixedScalars(nn.Module):
+            def forward(self) -> Tensor:
+                return torch.arange(0, 5, 0.5, dtype=torch.float32)
+
+        await validate_numerical_output(model=ArangeMixedScalars().eval())
+
         """Regression for ``replace_arange_start_step``: when ``end`` is
         SymInt-derived (carrying f32 element type from a sym_size cast)
         and ``start`` / ``step`` come in as scalar si32, ``coreai.range_``