diff --git a/tests/models/qwen3_vl_moe/test_modeling_qwen3_vl_moe.py b/tests/models/qwen3_vl_moe/test_modeling_qwen3_vl_moe.py
index 63e64de60b93..d61c0f7bd964 100644
--- a/tests/models/qwen3_vl_moe/test_modeling_qwen3_vl_moe.py
+++ b/tests/models/qwen3_vl_moe/test_modeling_qwen3_vl_moe.py
@@ -13,6 +13,7 @@
 # limitations under the License.
 """Testing suite for the PyTorch Qwen3VLMoe model."""
 
+import tempfile
 import unittest
 
 import pytest
@@ -27,6 +28,8 @@
 from transformers.models.qwen3_vl_moe.configuration_qwen3_vl_moe import Qwen3VLMoeTextConfig, Qwen3VLMoeVisionConfig
 from transformers.testing_utils import (
     Expectations,
+    backend_device_count,
+    get_cpu_ram_total_gib,
     require_flash_attn,
     require_torch,
     require_torch_accelerator,
@@ -301,6 +304,49 @@ def test_reverse_loading_mapping(self, check_keys_were_modified=False):
 class Qwen3VLMoeIntegrationTest(MemoryCleanupMixin, unittest.TestCase):
     maxDiff = None
 
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+        # Assigned here rather than in the class body: `MemoryCleanupMixin` snapshots the class body and
+        # protects it from teardown, so a checkpoint parked on a class-body `model` is never released.
+        cls.model = None
+        cls.offload_dir = None
+
+    @classmethod
+    def get_model(cls):
+        if cls.model is None:
+            cls.offload_dir = tempfile.TemporaryDirectory()
+            # device_map="auto" fills GPUs to ~100%, leaving no room for the transpose temporary
+            # buffer that fuses per-expert weight shards into a single gate_up_proj tensor during
+            # from_pretrained — causing CUDA OOM on multi-GPU runners. A 70% per-GPU max_memory cap
+            # reserves the headroom.
+            n = backend_device_count(torch_device)
+            if n > 0 and torch_device != "cpu":
+                torch_accel = getattr(torch, torch_device)
+                per_device = int(
+                    min(torch_accel.get_device_properties(i).total_memory for i in range(n)) * 0.70 / 1024**3
+                )
+                max_memory = dict.fromkeys(range(n), f"{per_device}GiB")
+                max_memory["cpu"] = (
+                    f"{int(get_cpu_ram_total_gib() * 0.9)}GiB"  # To avoid runner failing with exit code 137.
+                )
+            else:
+                max_memory = None
+            cls.model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
+                "Qwen/Qwen3-VL-30B-A3B-Instruct",
+                dtype="auto",
+                device_map="auto",
+                max_memory=max_memory,
+                offload_folder=cls.offload_dir.name,
+            )
+        return cls.model
+
+    @classmethod
+    def tearDownClass(cls):
+        if cls.offload_dir is not None:
+            cls.offload_dir.cleanup()
+        super().tearDownClass()
+
     def setUp(self):
         super().setUp()
 
@@ -351,9 +397,7 @@ def setUp(self):
 
     @slow
     def test_small_model_integration_test(self):
-        model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
-            "Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto"
-        )
+        model = self.get_model()
 
         inputs = self.processor.apply_chat_template(
             self.message, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
@@ -394,9 +438,7 @@ def test_small_model_integration_test(self):
 
     @slow
     def test_small_model_integration_test_batch(self):
-        model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
-            "Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto"
-        )
+        model = self.get_model()
         batch_messages = [self.message] * 2
         inputs = self.processor.apply_chat_template(
             batch_messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
@@ -462,9 +504,7 @@ def test_small_model_integration_test_with_video(self):
 
     @slow
     def test_small_model_integration_test_expand(self):
-        model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
-            "Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto"
-        )
+        model = self.get_model()
         inputs = self.processor.apply_chat_template(
             self.message, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
         ).to(torch_device)
@@ -490,9 +530,7 @@ def test_small_model_integration_test_expand(self):
 
     @slow
     def test_small_model_integration_test_expand_with_video(self):
-        model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
-            "Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto"
-        )
+        model = self.get_model()
         inputs = self.processor.apply_chat_template(
             self.message3, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
         ).to(torch_device)
@@ -511,9 +549,7 @@ def test_small_model_integration_test_expand_with_video(self):
 
     @slow
     def test_small_model_integration_test_batch_wo_image(self):
-        model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
-            "Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto"
-        )
+        model = self.get_model()
         message_wo_image = [
             {"role": "user", "content": [{"type": "text", "text": "Who are you?"}]},
         ]
@@ -541,9 +577,7 @@ def test_small_model_integration_test_batch_wo_image(self):
 
     @slow
     def test_small_model_integration_test_batch_different_resolutions(self):
-        model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
-            "Qwen/Qwen3-VL-30B-A3B-Instruct", dtype="auto", device_map="auto"
-        )
+        model = self.get_model()
         batched_messages = [self.message, self.message2]
         inputs = self.processor.apply_chat_template(
             batched_messages,