diff --git a/CHANGELOG.md b/CHANGELOG.md
index 88d92038..faaa52e1 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -38,6 +38,8 @@ Versioning](https://semver.org/spec/v2.0.0.html).
   and preventing memory allocations in the process.
 - Further optimized VST3 audio processing by preallocating small vectors for
   event and parameter change queues.
+- VST2 audio processing also received the same small vector optimization to get
+  rid of any last potential allocations during audio processing.
 - Changed the way mutual recursion in VST3 plugins on the plugin side works to
   counter any potential GUI related timing issues with VST3 plugins.
 - The deserialization part of yabridge's communication is now slightly faster by
diff --git a/src/common/serialization/vst2.h b/src/common/serialization/vst2.h
index 0b858ece..6afe9b7d 100644
--- a/src/common/serialization/vst2.h
+++ b/src/common/serialization/vst2.h
@@ -19,11 +19,13 @@
 #include <variant>
 
 #include <bitsery/ext/std_optional.h>
-#include "../bitsery/ext/in-place-variant.h"
 #include <bitsery/traits/array.h>
 #include <bitsery/traits/vector.h>
 #include <vestige/aeffectx.h>
+#include <boost/container/small_vector.hpp>
 
+#include "../bitsery/ext/in-place-variant.h"
+#include "../bitsery/traits/small-vector.h"
 #include "../utils.h"
 #include "../vst24.h"
 #include "common.h"
@@ -162,6 +164,9 @@ struct ChunkData {
  * Before serialization the events are read from a C-style array into a vector
  * using this class's constructor, and after deserializing the original struct
  * can be reconstructed using the `as_c_events()` method.
+ *
+ * Using preallocated small vectors here gets rid of all event related
+ * allocations in normal use cases.
  */
 class alignas(16) DynamicVstEvents {
    public:
@@ -177,9 +182,11 @@ class alignas(16) DynamicVstEvents {
     VstEvents& as_c_events();
 
     /**
-     * MIDI events are sent in batches.
+     * MIDI events are sent just before the audio processing call. Technically a
+     * host can call `effProcessEvents()` multiple times, but in practice this
+     * of course doesn't happen.
      */
-    std::vector<VstEvent> events;
+    boost::container::small_vector<VstEvent, 64> events;
 
     template <typename S>
     void serialize(S& s) {
@@ -194,14 +201,19 @@ class alignas(16) DynamicVstEvents {
      * `as_c_events()` method.
      *
      * The reason why this is necessary is because the `VstEvents` struct is
-     * actually a variable size object. In the definition in
+     * actuall variable size object. In the definition in
      * `vestige/aeffectx.h` the struct contains a single element `VstEvent`
      * pointer array, but the actual length of this array is
      * `VstEvents::numEvents`. Because there is no real limit on the number of
      * MIDI events the host can send at once we have to build this object on the
      * heap by hand.
      */
-    std::vector<uint8_t> vst_events_buffer;
+    boost::container::small_vector<
+        uint8_t,
+        sizeof(VstEvents) +
+            ((64 - 1) *
+             sizeof(VstEvent*))>  // NOLINT(bugprone-sizeof-expression)
+        vst_events_buffer;
 };
 
 /**
diff --git a/src/plugin/bridges/vst2.h b/src/plugin/bridges/vst2.h
index d46b2d1d..441d2595 100644
--- a/src/plugin/bridges/vst2.h
+++ b/src/plugin/bridges/vst2.h
@@ -185,10 +185,10 @@ class Vst2PluginBridge : PluginBridge<Vst2Sockets<std::jthread>> {
      * function. If they are sent during any other time or from another thread,
      * then the host will just discard them. Because we're receiving our host
      * callbacks on a separate thread, we have to temporarily store any events
-     * we receive so we can send them to the host at the end of
+     * we receive so we can send them to host on the audio thread at the end of
      * `process_replacing()`.
      */
-    std::vector<DynamicVstEvents> incoming_midi_events;
+    boost::container::small_vector<DynamicVstEvents, 4> incoming_midi_events;
     /**
      * Mutex for locking the above event queue, since recieving and processing
      * now happens in two different threads.
diff --git a/src/wine-host/bridges/vst2.h b/src/wine-host/bridges/vst2.h
index bb5ab6ed..5217f017 100644
--- a/src/wine-host/bridges/vst2.h
+++ b/src/wine-host/bridges/vst2.h
@@ -189,8 +189,13 @@ class Vst2Bridge : public HostBridge {
      * events they receive but some plugins such as Kontakt only store pointers
      * to these events, which means that the actual `VstEvent` objects must live
      * at least until the next audio buffer gets processed.
+     *
+     * Technically a host can send more than one of these at a time, but in
+     * practice every host will bundle all events in a single
+     * `effProcessEvents()` call.
      */
-    std::vector<DynamicVstEvents> next_audio_buffer_midi_events;
+    boost::container::small_vector<DynamicVstEvents, 4>
+        next_audio_buffer_midi_events;
     /**
      * Whether `next_audio_buffer_midi_events` should be cleared before
      * inserting new events.