diff --git a/tests/hijax_test.py b/tests/hijax_test.py
index b0cbcf4541b4..311c0f069fa2 100644
--- a/tests/hijax_test.py
+++ b/tests/hijax_test.py
@@ -535,6 +535,66 @@ def f(x):
     self.assertAllClose(jax.vmap(f)(xs), xs**3)
     self.assertEqual(jax.grad(f)(2.0), 12.0)
 
+  @config.numpy_dtype_promotion('standard')
+  def test_newstyle_hiprimitive_qarray(self):
+
+    @dataclass(frozen=True)  # not NamedTuple, which is a pytree
+    class QArray:
+      qvalue: jax.Array
+      scale: jax.Array
+
+    @dataclass(frozen=True)
+    class QArrayTy(HiType):
+      shape: tuple[int, int]
+
+      def to_tangent_aval(self):
+        return ShapedArray(self.shape, jnp.dtype('float32'))
+
+    register_hitype(QArray, lambda q: QArrayTy(q.qvalue.shape))
+
+    def q(x):
+      return Q(jax.typeof(x))(x)
+
+    def dq(qx):
+      return DQ(jax.typeof(qx))(qx)
+
+    class Q(NewstyleHiPrimitive):
+      def __init__(self, unquantized_aval):
+        if unquantized_aval.dtype != jnp.dtype('float32'): raise TypeError
+        quantized_aval = QArrayTy(unquantized_aval.shape)
+        super().__init__((unquantized_aval,), quantized_aval)
+
+      def expand(self, x):
+        scale = jnp.max(jnp.abs(x)) / 127
+        qvalue = jnp.round(x / scale).astype(jnp.int8)
+        return QArray(qvalue, scale)
+
+      def vjp_fwd(self, x):
+        return self(x), None
+
+      def vjp_bwd(self, _, g):
+        return g,
+
+    class DQ(NewstyleHiPrimitive):
+      def __init__(self, quantized_aval):
+        unquantized_aval = ShapedArray(quantized_aval.shape, jnp.dtype('float32'))
+        super().__init__((quantized_aval,), unquantized_aval)
+
+      def expand(self, qx):
+        return qx.qvalue * qx.scale
+
+      def vjp_fwd(self, qx):
+        return self(qx), None
+
+      def vjp_bwd(self, _, g):
+        return g,
+
+    def f(x):
+      return jnp.sum(dq(q(x)))
+
+    x = jax.random.normal(jax.random.key(0), (3, 3), dtype='float32')
+    g = jax.grad(f)(x)
+
 
 class BoxTest(jtu.JaxTestCase):