@@ -3988,6 +3988,12 @@ def _deploy_core_endpoint(self, **kwargs):
                     "Fast Model Loading. Configure by setting `num_cpus` to 0 in `resources`."
                 )
 
+        if serverless_inference_config is None:
+            inference_ami_version = self._resolve_inference_ami_version(
+                instance_type=instance_type,
+                inference_ami_version=inference_ami_version,
+            )
+
         if endpoint_type == EndpointType.INFERENCE_COMPONENT_BASED:
             if update_endpoint:
                 raise ValueError(
@@ -6053,6 +6059,9 @@ def deploy(
         if not hasattr(self, "built_model") and not hasattr(self, "_deployables"):
             raise ValueError("Model needs to be built before deploying")
 
+        if instance_type:
+            self.instance_type = instance_type
+
         # Inference component deployments manage their own reuse by IC name
         # (create vs. in-place update in _deploy_for_ic). The endpoint-return
         # reuse gate must not intercept them, or an intended IC create/update
@@ -6113,11 +6122,6 @@ def deploy(
             if not self.instance_type and not instance_type:
                 self.instance_type = self._fetch_default_instance_type_for_custom_model()
 
-            # Ensure self.instance_type reflects the caller's intent so the
-            # endpoint config creation in _deploy_model_customization picks it up.
-            if instance_type:
-                self.instance_type = instance_type
-
             # Pass inference_config if it's ResourceRequirements
             inference_config_param = None
             if isinstance(inference_config, ResourceRequirements):
 
@@ -21,7 +21,7 @@
 
 import os
 from pathlib import Path
-from typing import Union
+from typing import Optional, Union
 
 # SageMaker core imports
 from sagemaker.core.resources import Model, Endpoint
@@ -76,6 +76,22 @@
 MODEL_SERVER_WORKERS_PARAM_NAME = "sagemaker_model_server_workers"
 SAGEMAKER_REGION_PARAM_NAME = "sagemaker_region"
 SAGEMAKER_OUTPUT_LOCATION = "sagemaker_s3_output"
+# Intentionally allowlist only instance families documented for the GPU 4.1 AMI.
+# Add new families only after SageMaker documents support for them.
+_CUDA_13_INFERENCE_AMI = "al2023-ami-sagemaker-inference-gpu-4-1"
+_CUDA_13_INFERENCE_AMI_COMPATIBLE_FAMILIES = frozenset(
+    {
+        "ml.g4dn",
+        "ml.g5",
+        "ml.g6",
+        "ml.g6e",
+        "ml.p4d",
+        "ml.p4de",
+        "ml.p5",
+        "ml.p5e",
+        "ml.p5en",
+    }
+)
 
 
 class _ModelBuilderServers(object):
@@ -327,15 +343,37 @@ def _build_for_tgi(self) -> Model:
 
         return model
 
+    def _resolve_inference_ami_version(
+        self,
+        instance_type: Optional[str],
+        inference_ami_version: Optional[str],
+    ) -> Optional[str]:
+        """Resolve a CUDA-compatible inference AMI without overriding the caller."""
+        if inference_ami_version is not None:
+            return inference_ami_version
+        if not isinstance(self.image_uri, str) or not isinstance(instance_type, str):
+            return inference_ami_version
+
+        # DLC CUDA versions are hyphen-delimited tag tokens. Splitting also
+        # recognizes a tag ending in ``-cu130`` without requiring a suffix.
+        image_tag = self.image_uri.rpartition(":")[2]
+        if "cu130" not in image_tag.split("-"):
+            return inference_ami_version
+
+        instance_family = instance_type.rsplit(".", 1)[0]
+        if instance_family not in _CUDA_13_INFERENCE_AMI_COMPATIBLE_FAMILIES:
+            return inference_ami_version
+
+        return _CUDA_13_INFERENCE_AMI
+
     def _build_for_hf_server(self, model_server: ModelServer) -> Model:
         """Build a HuggingFace model for a hub-download serving container.
 
         Generic build path shared by the vLLM, SGLang, and vLLM-omni servers. It
-        configures the container to pull the model directly from the HuggingFace Hub
-        (HF_MODEL_ID), resolves the appropriate DLC image via _auto_detect_image_uri,
-        and prepares the model for the selected mode. Server-specific tuning (sharding,
-        tensor parallelism, MAX_* limits, etc.) is intentionally left to the container
-        defaults and will be added in a follow-up.
+        configures the container to pull the model directly from the HuggingFace Hub,
+        resolves the appropriate DLC image via _auto_detect_image_uri, and prepares the
+        model for the selected mode. Server-specific tuning (sharding, tensor parallelism,
+        MAX_* limits, etc.) remains configurable through environment variables.
 
         Args:
             model_server: The HuggingFace serving model server to build for
@@ -358,8 +396,7 @@ def _build_for_hf_server(self, model_server: ModelServer) -> Model:
         _create_dir_structure(self.model_path)
 
         if isinstance(self.model, str) and not self._is_jumpstart_model_id():
-            # These containers download the model directly from the HuggingFace Hub
-            # Todo: missing something?
+            # These containers download the model directly from the HuggingFace Hub.
             self.env_vars.setdefault("HF_MODEL_ID", self.model)
 
             if self.env_vars.get("HUGGING_FACE_HUB_TOKEN"):
 
@@ -1,4 +1,4 @@
-"""Unit tests: HF_MODEL_ID is not overwritten when user provides it."""
+"""Unit tests for Hugging Face ModelBuilder server configuration."""
 from __future__ import annotations
 
 from typing import Dict, List, Optional
@@ -292,6 +292,102 @@ def test_sets_default_hf_model_id_when_not_provided(
     )
 
 
+class TestCuda13AmiResolution:
+    """CUDA 13 serving images select compatible host drivers at deployment."""
+
+    def test_selects_cuda_13_ami_for_vllm_omni(
+        self,
+        mock_builder: MagicMock,
+    ) -> None:
+        """The deploy-time G5 instance overrides the M5 instance used at build time."""
+        builder = mock_builder
+        builder.instance_type = "ml.m5.large"
+        builder.image_uri = (
+            "123456789012.dkr.ecr.us-west-2.amazonaws.com/"
+            "huggingface-vllm-omni:0.20.0-gpu-py312-cu130-amzn2023"
+        )
+
+        result = _ModelBuilderServers._resolve_inference_ami_version(
+            builder,
+            instance_type="ml.g5.2xlarge",
+            inference_ami_version=None,
+        )
+
+        assert result == "al2023-ami-sagemaker-inference-gpu-4-1"
+
+    def test_recognizes_image_tag_ending_in_cu130(
+        self,
+        mock_builder: MagicMock,
+    ) -> None:
+        """A cu130 token remains detectable without a trailing tag component."""
+        builder = mock_builder
+        builder.image_uri = (
+            "123456789012.dkr.ecr.us-west-2.amazonaws.com/"
+            "huggingface-vllm:0.21.0-gpu-py312-cu130"
+        )
+
+        result = _ModelBuilderServers._resolve_inference_ami_version(
+            builder,
+            instance_type="ml.g5.2xlarge",
+            inference_ami_version=None,
+        )
+
+        assert result == "al2023-ami-sagemaker-inference-gpu-4-1"
+
+    @pytest.mark.parametrize(
+        "image_uri, instance_type",
+        [
+            (
+                "123456789012.dkr.ecr.us-west-2.amazonaws.com/"
+                "huggingface-vllm:0.17.0-gpu-py312-cu129-ubuntu22.04",
+                "ml.g5.2xlarge",
+            ),
+            (
+                "123456789012.dkr.ecr.us-west-2.amazonaws.com/"
+                "huggingface-vllm:0.21.0-gpu-py312-cu130-ubuntu22.04",
+                "ml.m5.large",
+            ),
+        ],
+        ids=["non-cu130-image", "unsupported-instance-family"],
+    )
+    def test_leaves_ami_unset_outside_compatibility_gate(
+        self,
+        mock_builder: MagicMock,
+        image_uri: str,
+        instance_type: str,
+    ) -> None:
+        """Images and instances outside the compatibility gate retain the default."""
+        builder = mock_builder
+        builder.image_uri = image_uri
+
+        result = _ModelBuilderServers._resolve_inference_ami_version(
+            builder,
+            instance_type=instance_type,
+            inference_ami_version=None,
+        )
+
+        assert result is None
+
+    def test_preserves_explicit_inference_ami(
+        self,
+        mock_builder: MagicMock,
+    ) -> None:
+        """An explicit deployment AMI takes precedence over automatic selection."""
+        builder = mock_builder
+        builder.image_uri = (
+            "123456789012.dkr.ecr.us-west-2.amazonaws.com/"
+            "huggingface-vllm:0.21.0-gpu-py312-cu130-ubuntu22.04"
+        )
+
+        result = _ModelBuilderServers._resolve_inference_ami_version(
+            builder,
+            instance_type="ml.g5.2xlarge",
+            inference_ami_version="custom-inference-ami",
+        )
+
+        assert result == "custom-inference-ami"
+
+
 # -----------------------------------------------------------
 # Transformers (MMS) — needs extra patches
 # -----------------------------------------------------------
 
@@ -198,12 +198,13 @@ def test_deploy_generates_unique_endpoint_name(self, mock_deploy):
         )
         builder.built_model = Mock(spec=Model)
 
-        result = builder.deploy(instance_type="ml.m5.large")
-        
+        result = builder.deploy(instance_type="ml.g5.xlarge")
+
         self.assertIsNotNone(result)
-        # Verify endpoint name was generated (contains uuid)
+        # Verify endpoint name generation and the deploy-time instance override.
         self.assertIn("endpoint-", builder.endpoint_name)
-        mock_deploy.assert_called_once()
+        self.assertEqual(builder.instance_type, "ml.g5.xlarge")
+        self.assertEqual(mock_deploy.call_args.kwargs["instance_type"], "ml.g5.xlarge")
 
     @patch('sagemaker.serve.model_builder.ModelBuilder._deploy')
     def test_deploy_with_serverless_config(self, mock_deploy):