diff --git a/include/rendering/m2_renderer.hpp b/include/rendering/m2_renderer.hpp
index de4fbf84..5e402fe6 100644
--- a/include/rendering/m2_renderer.hpp
+++ b/include/rendering/m2_renderer.hpp
@@ -357,6 +357,9 @@ public:
     void removeInstances(const std::vector<uint32_t>& instanceIds);
     void setSkipCollision(uint32_t instanceId, bool skip);
     void clear();
+    /** Drop all instances but keep models in GPU memory. Cheap path for the
+     *  editor's rebuild loop where the same model is re-instanced repeatedly. */
+    void clearInstances();
     void cleanupUnusedModels();
 
     bool checkCollision(const glm::vec3& from, const glm::vec3& to,
diff --git a/src/rendering/m2_renderer_instance.cpp b/src/rendering/m2_renderer_instance.cpp
index 269fab62..caa16f3c 100644
--- a/src/rendering/m2_renderer_instance.cpp
+++ b/src/rendering/m2_renderer_instance.cpp
@@ -377,6 +377,22 @@ void M2Renderer::clear() {
     textureBudgetRejectWarnings_ = 0;
 }
 
+void M2Renderer::clearInstances() {
+    if (vkCtx_) vkDeviceWaitIdle(vkCtx_->getDevice());
+    for (auto& inst : instances) destroyInstanceBones(inst);
+    instances.clear();
+    spatialGrid.clear();
+    instanceIndexById.clear();
+    instanceDedupMap_.clear();
+    smokeInstanceIndices_.clear();
+    portalInstanceIndices_.clear();
+    animatedInstanceIndices_.clear();
+    particleOnlyInstanceIndices_.clear();
+    particleInstanceIndices_.clear();
+    smokeParticles.clear();
+    smokeEmitAccum = 0.0f;
+}
+
 void M2Renderer::setCollisionFocus(const glm::vec3& worldPos, float radius) {
     collisionFocusEnabled = (radius > 0.0f);
     collisionFocusPos = worldPos;
diff --git a/tools/editor/editor_viewport.cpp b/tools/editor/editor_viewport.cpp
index f56e04ba..fea9610b 100644
--- a/tools/editor/editor_viewport.cpp
+++ b/tools/editor/editor_viewport.cpp
@@ -80,6 +80,13 @@ bool EditorViewport::loadTerrain(const pipeline::TerrainMesh& mesh,
 
 void EditorViewport::clearTerrain() {
     if (terrainRenderer_) terrainRenderer_->clear();
+    // Loading a different zone invalidates the cached models; flush them so
+    // their slots can be reused without leaking GPU memory across zones.
+    persistentM2ModelIds_.clear();
+    persistentWMOModelIds_.clear();
+    nextPersistentModelId_ = 1;
+    if (m2Renderer_) m2Renderer_->clear();
+    if (wmoRenderer_) wmoRenderer_->clearAll();
 }
 
 void EditorViewport::updateWater(const pipeline::ADTTerrain& terrain, int tileX, int tileY) {
@@ -106,12 +113,16 @@ void EditorViewport::clearObjects() {
     ghostModelId_ = 0;
     ghostModelPath_.clear();
 
+    // Drop instances but keep models cached on the GPU. The editor's rebuild
+    // path destroys-and-recreates instances every time the placement set
+    // changes; preserving model GPU buffers makes that path much cheaper for
+    // large NPC populations using shared models.
     if (m2Renderer_) {
         vkCtx_->waitAllUploads();
-        m2Renderer_->clear();
+        m2Renderer_->clearInstances();
     }
     if (wmoRenderer_) {
-        wmoRenderer_->clearAll();
+        wmoRenderer_->clearInstances();
     }
 }
 
@@ -120,9 +131,11 @@ void EditorViewport::rebuildObjects(const std::vector<PlacedObject>& objects,
     clearObjects();
     if (objects.empty() && npcs.empty()) return;
 
-    // Don't call beginUploadBatch here — loadModel starts its own batch
-    uint32_t nextModelId = 1;
-    std::unordered_map<std::string, uint32_t> m2ModelIds, wmoModelIds;
+    // Don't call beginUploadBatch here — loadModel starts its own batch.
+    // Use the persistent model-id maps so models stay cached across rebuilds.
+    auto& m2ModelIds = persistentM2ModelIds_;
+    auto& wmoModelIds = persistentWMOModelIds_;
+    uint32_t& nextModelId = nextPersistentModelId_;
 
     for (const auto& obj : objects) {
         if (obj.type == PlaceableType::M2 && m2Renderer_) {
diff --git a/tools/editor/editor_viewport.hpp b/tools/editor/editor_viewport.hpp
index d173f5e6..307dab18 100644
--- a/tools/editor/editor_viewport.hpp
+++ b/tools/editor/editor_viewport.hpp
@@ -13,6 +13,8 @@
 #include <vulkan/vulkan.h>
 #include <vk_mem_alloc.h>
 #include <memory>
+#include <string>
+#include <unordered_map>
 
 namespace wowee {
 namespace pipeline { class AssetManager; }
@@ -114,6 +116,12 @@ private:
     float fogNear_ = 5000.0f, fogFar_ = 10000.0f;
     float timeOfDay_ = 12.0f;
 
+    // Persistent path -> renderer model ID maps. Keeping these across rebuilds
+    // lets the renderer skip re-uploading models that are still in its cache.
+    std::unordered_map<std::string, uint32_t> persistentM2ModelIds_;
+    std::unordered_map<std::string, uint32_t> persistentWMOModelIds_;
+    uint32_t nextPersistentModelId_ = 1;
+
     // Ghost preview state
     std::string ghostModelPath_;
     uint32_t ghostModelId_ = 0;