diff --git a/doc/source/whatsnew/v3.0.0.rst b/doc/source/whatsnew/v3.0.0.rst
index 0f7be8cfbcb68..b121c5da435b6 100644
--- a/doc/source/whatsnew/v3.0.0.rst
+++ b/doc/source/whatsnew/v3.0.0.rst
@@ -1127,6 +1127,7 @@ Interval
 - :meth:`Index.is_monotonic_decreasing`, :meth:`Index.is_monotonic_increasing`, and :meth:`Index.is_unique` could incorrectly be ``False`` for an ``Index`` created from a slice of another ``Index``. (:issue:`57911`)
 - Bug in :class:`Index`, :class:`Series`, :class:`DataFrame` constructors when given a sequence of :class:`Interval` subclass objects casting them to :class:`Interval` (:issue:`46945`)
 - Bug in :func:`interval_range` where start and end numeric types were always cast to 64 bit (:issue:`57268`)
+- Bug in :func:`pandas.interval_range` incorrectly inferring ``int64`` dtype when ``np.float32`` and ``int`` are used for ``start`` and ``freq`` (:issue:`58964`)
 - Bug in :meth:`IntervalIndex.get_indexer` and :meth:`IntervalIndex.drop` when one of the sides of the index is non-unique (:issue:`52245`)
 - Construction of :class:`IntervalArray` and :class:`IntervalIndex` from arrays with mismatched signed/unsigned integer dtypes (e.g., ``int64`` and ``uint64``) now raises a :exc:`TypeError` instead of proceeding silently. (:issue:`55715`)
 
diff --git a/pandas/core/indexes/interval.py b/pandas/core/indexes/interval.py
index a4a6230337add..37b1838665ee9 100644
--- a/pandas/core/indexes/interval.py
+++ b/pandas/core/indexes/interval.py
@@ -1420,17 +1420,17 @@ def interval_range(
         dtype: np.dtype = np.dtype("int64")
         if com.all_not_none(start, end, freq):
             if (
-                isinstance(start, (float, np.float16))
-                or isinstance(end, (float, np.float16))
-                or isinstance(freq, (float, np.float16))
-            ):
-                dtype = np.dtype("float64")
-            elif (
                 isinstance(start, (np.integer, np.floating))
                 and isinstance(end, (np.integer, np.floating))
                 and start.dtype == end.dtype
             ):
                 dtype = start.dtype
+            elif (
+                isinstance(start, (float, np.floating))
+                or isinstance(end, (float, np.floating))
+                or isinstance(freq, (float, np.floating))
+            ):
+                dtype = np.dtype("float64")
             # 0.1 ensures we capture end
             breaks = np.arange(start, end + (freq * 0.1), freq)
             breaks = maybe_downcast_numeric(breaks, dtype)
diff --git a/pandas/tests/indexes/interval/test_interval_range.py b/pandas/tests/indexes/interval/test_interval_range.py
index 5252b85ad8d0e..66c0111edb16e 100644
--- a/pandas/tests/indexes/interval/test_interval_range.py
+++ b/pandas/tests/indexes/interval/test_interval_range.py
@@ -380,3 +380,11 @@ def test_float_freq(self):
         result = interval_range(0, 1, freq=0.6)
         expected = IntervalIndex.from_breaks([0, 0.6])
         tm.assert_index_equal(result, expected)
+
+    def test_interval_range_float32_start_int_freq(self):
+        # GH 58964
+        result = interval_range(start=np.float32(0), end=2, freq=1)
+        expected = IntervalIndex.from_tuples(
+            [(0.0, 1.0), (1.0, 2.0)], dtype="interval[float64, right]"
+        )
+        tm.assert_index_equal(result, expected)