diff --git a/src/nfa/shengcompile.cpp b/src/nfa/shengcompile.cpp
index 8b939973..f968bf59 100644
--- a/src/nfa/shengcompile.cpp
+++ b/src/nfa/shengcompile.cpp
@@ -643,19 +643,40 @@ bytecode_ptr<NFA> shengCompile(raw_dfa &raw, const CompileContext &cc,
     DEBUG_PRINTF("This DFA %s die so effective number of states is %zu\n",
                  info.can_die ? "can" : "cannot", info.size());
     if (info.size() > 16) {
-#if defined(HAVE_AVX512VBMI)
-        if (info.size() > 32) {
-            DEBUG_PRINTF("Too many states\n");
-            return nullptr;
-        }
-        return shengCompile_int<sheng32>(raw, cc, accel_states, strat, info);
-#else
         DEBUG_PRINTF("Too many states\n");
         return nullptr;
-#endif
     }
 
     return shengCompile_int<sheng>(raw, cc, accel_states, strat, info);
 }
 
+#if defined(HAVE_AVX512VBMI)
+bytecode_ptr<NFA> sheng32Compile(raw_dfa &raw, const CompileContext &cc,
+                                 const ReportManager &rm, bool only_accel_init,
+                                 set<dstate_id_t> *accel_states) {
+    if (!cc.grey.allowSheng) {
+        DEBUG_PRINTF("Sheng is not allowed!\n");
+        return nullptr;
+    }
+
+    sheng_build_strat strat(raw, rm, only_accel_init);
+    dfa_info info(strat);
+
+    DEBUG_PRINTF("Trying to compile a %zu state Sheng\n", raw.states.size());
+
+    DEBUG_PRINTF("Anchored start state id: %u, floating start state id: %u\n",
+                 raw.start_anchored, raw.start_floating);
+
+    DEBUG_PRINTF("This DFA %s die so effective number of states is %zu\n",
+                 info.can_die ? "can" : "cannot", info.size());
+    assert(info.size() > 16);
+    if (info.size() > 32) {
+        DEBUG_PRINTF("Too many states\n");
+        return nullptr;
+    }
+
+    return shengCompile_int<sheng32>(raw, cc, accel_states, strat, info);
+}
+#endif
+
 } // namespace ue2
diff --git a/src/nfa/shengcompile.h b/src/nfa/shengcompile.h
index d795b362..b36e27be 100644
--- a/src/nfa/shengcompile.h
+++ b/src/nfa/shengcompile.h
@@ -71,6 +71,12 @@ bytecode_ptr<NFA> shengCompile(raw_dfa &raw, const CompileContext &cc,
                                const ReportManager &rm, bool only_accel_init,
                                std::set<dstate_id_t> *accel_states = nullptr);
 
+#if defined(HAVE_AVX512VBMI)
+bytecode_ptr<NFA> sheng32Compile(raw_dfa &raw, const CompileContext &cc,
+                                 const ReportManager &rm, bool only_accel_init,
+                                 std::set<dstate_id_t> *accel_states = nullptr);
+#endif
+
 struct sheng_escape_info {
     CharReach outs;
     CharReach outs2_single;
diff --git a/src/rose/rose_build_bytecode.cpp b/src/rose/rose_build_bytecode.cpp
index 5cbb5c84..8e1d7095 100644
--- a/src/rose/rose_build_bytecode.cpp
+++ b/src/rose/rose_build_bytecode.cpp
@@ -632,6 +632,11 @@ bytecode_ptr<NFA> getDfa(raw_dfa &rdfa, bool is_transient,
          * bytecode and that they are usually run on small blocks */
         dfa = mcshengCompile(rdfa, cc, rm);
     }
+#if defined(HAVE_AVX512VBMI)
+    if (!dfa) {
+        dfa = sheng32Compile(rdfa, cc, rm, false);
+    }
+#endif
     if (!dfa) {
         // Sheng wasn't successful, so unleash McClellan!
         dfa = mcclellanCompile(rdfa, cc, rm, false);
diff --git a/src/smallwrite/smallwrite_build.cpp b/src/smallwrite/smallwrite_build.cpp
index 345edfe9..909fdcb3 100644
--- a/src/smallwrite/smallwrite_build.cpp
+++ b/src/smallwrite/smallwrite_build.cpp
@@ -793,6 +793,11 @@ bytecode_ptr<NFA> getDfa(raw_dfa &rdfa, const CompileContext &cc,
     bytecode_ptr<NFA> dfa = nullptr;
     if (cc.grey.allowSmallWriteSheng) {
         dfa = shengCompile(rdfa, cc, rm, only_accel_init, &accel_states);
+#if defined(HAVE_AVX512VBMI)
+        if (!dfa) {
+            dfa = sheng32Compile(rdfa, cc, rm, only_accel_init, &accel_states);
+        }
+#endif
     }
     if (!dfa) {
         dfa = mcclellanCompile(rdfa, cc, rm, only_accel_init,