diff --git a/py/mpz.c b/py/mpz.c
index 80473df882d8815dca10e5af72011ce1ab05d7e8..f58e262e2358110e9e2586d953a782ea6cb40079 100644
--- a/py/mpz.c
+++ b/py/mpz.c
@@ -946,10 +946,6 @@ bool mpz_is_pos(const mpz_t *z) {
     return z->len > 0 && z->neg == 0;
 }
 
-bool mpz_is_neg(const mpz_t *z) {
-    return z->len > 0 && z->neg != 0;
-}
-
 bool mpz_is_odd(const mpz_t *z) {
     return z->len > 0 && (z->dig[0] & 1) != 0;
 }
diff --git a/py/mpz.h b/py/mpz.h
index 2ff404ffa839f65117af1b150139e54fb953f93f..55967cc4c30188a5020b15781244ce7132ab10d3 100644
--- a/py/mpz.h
+++ b/py/mpz.h
@@ -112,6 +112,7 @@ size_t mpz_set_from_str(mpz_t *z, const char *str, size_t len, bool neg, unsigne
 void mpz_set_from_bytes(mpz_t *z, bool big_endian, size_t len, const byte *buf);
 
 static inline bool mpz_is_zero(const mpz_t *z) { return z->len == 0; }
+static inline bool mpz_is_neg(const mpz_t *z) { return z->len != 0 && z->neg != 0; }
 int mpz_cmp(const mpz_t *lhs, const mpz_t *rhs);
 
 void mpz_abs_inpl(mpz_t *dest, const mpz_t *z);
diff --git a/py/objint_longlong.c b/py/objint_longlong.c
index eb80407bcbc1deee44d87ac51d2e79c6b8d05437..1d184a7dc25c06d80b0eec01829d2da37787f332 100644
--- a/py/objint_longlong.c
+++ b/py/objint_longlong.c
@@ -191,6 +191,13 @@ mp_obj_t mp_obj_int_binary_op(mp_uint_t op, mp_obj_t lhs_in, mp_obj_t rhs_in) {
 
         case MP_BINARY_OP_POWER:
         case MP_BINARY_OP_INPLACE_POWER: {
+            if (rhs_val < 0) {
+                #if MICROPY_PY_BUILTINS_FLOAT
+                return mp_obj_float_binary_op(op, lhs_val, rhs_in);
+                #else
+                mp_raise_ValueError("negative power with no float support");
+                #endif
+            }
             long long ans = 1;
             while (rhs_val > 0) {
                 if (rhs_val & 1) {
diff --git a/py/objint_mpz.c b/py/objint_mpz.c
index d818b6f4078168f9cc9bb6a498499985cbdac762..26492aab48e747025f45552059fbd6a6bdf1c687 100644
--- a/py/objint_mpz.c
+++ b/py/objint_mpz.c
@@ -290,6 +290,13 @@ mp_obj_t mp_obj_int_binary_op(mp_uint_t op, mp_obj_t lhs_in, mp_obj_t rhs_in) {
 
             case MP_BINARY_OP_POWER:
             case MP_BINARY_OP_INPLACE_POWER:
+                if (mpz_is_neg(zrhs)) {
+                    #if MICROPY_PY_BUILTINS_FLOAT
+                    return mp_obj_float_binary_op(op, mpz_as_float(zlhs), rhs_in);
+                    #else
+                    mp_raise_ValueError("negative power with no float support");
+                    #endif
+                }
                 mpz_pow_inpl(&res->mpz, zlhs, zrhs);
                 break;
 
diff --git a/tests/float/int_big_float.py b/tests/float/int_big_float.py
index b1a26ca73d51a2890185641832b647c723094049..0bd166218664a82fb2c2cee982b7876abdefbc92 100644
--- a/tests/float/int_big_float.py
+++ b/tests/float/int_big_float.py
@@ -18,6 +18,10 @@ print("%.5g" % (i / 1.2))
 # this should delegate to complex
 print("%.5g" % (i * 1.2j).imag)
 
+# negative power should produce float
+print("%.5g" % (i ** -1))
+print("%.5g" % ((2 + i - i) ** -3))
+
 try:
     i / 0
 except ZeroDivisionError: