diff --git a/examples/recipes/usyd-community_vitpose-plus-huge/cpu/cpu/keypoint-detection_fp16_config.json b/examples/recipes/usyd-community_vitpose-plus-huge/cpu/cpu/keypoint-detection_fp16_config.json
new file mode 100644
index 000000000..7713b8e7d
--- /dev/null
+++ b/examples/recipes/usyd-community_vitpose-plus-huge/cpu/cpu/keypoint-detection_fp16_config.json
@@ -0,0 +1,61 @@
+{
+  "export": {
+    "opset_version": 17,
+    "batch_size": 1,
+    "export_params": true,
+    "do_constant_folding": true,
+    "verbose": false,
+    "dynamo": false,
+    "enable_hierarchy_tags": true,
+    "clean_onnx": false,
+    "hierarchy_tag_format": "full",
+    "input_tensors": [
+      {
+        "name": "pixel_values",
+        "dtype": "float32",
+        "shape": [
+          1,
+          3,
+          256,
+          192
+        ],
+        "value_range": [
+          0,
+          1
+        ]
+      }
+    ],
+    "output_tensors": [
+      {
+        "name": "heatmaps"
+      }
+    ]
+  },
+  "optim": {},
+  "quant": null,
+  "compile": null,
+  "loader": {
+    "task": "keypoint-detection",
+    "model_class": "VitPoseForPoseEstimation",
+    "model_type": "vitpose"
+  },
+  "eval": {
+    "task": "keypoint-detection",
+    "dataset": {
+      "path": "~/.cache/winml/datasets/coco_keypoints_val2017",
+      "build_script": "scripts/build_coco_keypoints.py",
+      "split": "validation",
+      "samples": 100,
+      "shuffle": true,
+      "seed": 42,
+      "columns_mapping": {
+        "input_column": "image",
+        "annotation_column": "objects",
+        "keypoints_key": "keypoints",
+        "bbox_key": "bbox",
+        "area_key": "area",
+        "box_format": "xywh"
+      }
+    }
+  }
+}
diff --git a/src/winml/modelkit/quant/fp16.py b/src/winml/modelkit/quant/fp16.py
index b3a6fe3a3..e35a3e35c 100644
--- a/src/winml/modelkit/quant/fp16.py
+++ b/src/winml/modelkit/quant/fp16.py
@@ -64,10 +64,33 @@ def convert_to_fp16(
     if op_block_list:
         logger.info("  Keeping ops in FP32: %s", op_block_list)
 
+    # ORT's converter runs ONNX shape inference by default, which serializes the
+    # model to a single protobuf message. That serialization fails for models
+    # whose in-memory proto exceeds protobuf's hard 2 GB message limit
+    # (google.protobuf.message.EncodeError: Failed to serialize proto), e.g.
+    # ViT-Huge backbones. Disable shape inference for such models — the FP16
+    # cast itself does not require inferred shapes.
+    protobuf_limit = 2 * 1024**3
+    shape_infer_margin = int(protobuf_limit * 0.9)  # stay safely under 2 GB
+    try:
+        # ByteSize() itself raises once the proto exceeds the 2 GB limit; that
+        # is already a definitive "too large" signal.
+        approx_size = model.ByteSize()
+    except Exception:
+        approx_size = protobuf_limit
+    disable_shape_infer = approx_size >= shape_infer_margin
+    if disable_shape_infer:
+        logger.info(
+            "  Model is large (~%.2f GB); disabling shape inference during FP16 "
+            "conversion to avoid the 2 GB protobuf serialization limit.",
+            approx_size / 1024**3,
+        )
+
     converted: ModelProto = convert_float_to_float16(
         model,
         keep_io_types=keep_io_types,
         op_block_list=op_block_list,
+        disable_shape_infer=disable_shape_infer,
     )
 
     # ORT's converter appends Cast nodes at the end of the node list (for