diff --git a/optimum/exporters/openvino/base.py b/optimum/exporters/openvino/base.py
index fe3ec819ae..86e9b258b6 100644
--- a/optimum/exporters/openvino/base.py
+++ b/optimum/exporters/openvino/base.py
@@ -352,9 +352,11 @@ def generate_dummy_inputs(self, framework: str = "pt", **kwargs):
             and self.PAD_ATTENTION_MASK_TO_PAST
             and self.use_cache_branch is not False
             and "attention_mask" in dummy_inputs
-            and self.task == "text-generation"
+            and self.task in ("text-generation", "image-text-to-text")
         ):
-            seq_len = dummy_inputs["input_ids"].shape[1]
+            # VLM Eagle3 uses inputs_embeds instead of input_ids
+            main_input = dummy_inputs.get("input_ids", dummy_inputs.get("inputs_embeds"))
+            seq_len = main_input.shape[1]
             past_seq_len = dummy_inputs["past_key_values"][0][1].shape[-2]
             dummy_inputs["attention_mask"] = DummyInputGenerator.pad_input_on_dim(
                 dummy_inputs["attention_mask"], desired_length=past_seq_len + seq_len, dim=1
diff --git a/optimum/intel/openvino/modeling_visual_language.py b/optimum/intel/openvino/modeling_visual_language.py
index 1fe784455c..9e335e06c7 100644
--- a/optimum/intel/openvino/modeling_visual_language.py
+++ b/optimum/intel/openvino/modeling_visual_language.py
@@ -895,6 +895,19 @@ def _from_pretrained(
             trust_remote_code (`bool`, *optional*, defaults to `False`):
                 Whether to trust remote code when loading model tokenizer/processor during quantization.
         """
+        if config.model_type not in MODEL_TYPE_TO_CLS_MAPPING:
+            archs = getattr(config, "architectures", None) or []
+            if archs and "eagle3" in archs[0].lower():
+                raise ValueError(
+                    f"Model with architecture '{archs[0]}' (model_type='{config.model_type}') is a standalone "
+                    "Eagle3 speculative-decoding draft model, not a multi-component VLM, even though its "
+                    "config declares a VLM-oriented `modal_type`/`target_model_type`. Please load it with "
+                    "`OVModelForCausalLM` instead of `OVModelForVisualCausalLM`."
+                )
+            raise ValueError(
+                f"Unsupported model_type '{config.model_type}' for `OVModelForVisualCausalLM`. Supported "
+                f"model types are: {sorted(MODEL_TYPE_TO_CLS_MAPPING)}."
+            )
         model_cls = MODEL_TYPE_TO_CLS_MAPPING[config.model_type]
         model_file_names = model_cls._all_ov_model_paths.copy()
         for k in tuple(model_file_names):
diff --git a/tests/openvino/test_exporters_cli.py b/tests/openvino/test_exporters_cli.py
index c57789a4c5..d082354750 100644
--- a/tests/openvino/test_exporters_cli.py
+++ b/tests/openvino/test_exporters_cli.py
@@ -1114,6 +1114,32 @@ def test_exporters_cli_int8(self, task: str, model_type: str):
                 del expected_int8["decoder_with_past"]
             check_compression_state_per_model(self, model.ov_models, expected_int8)
 
+    @parameterized.expand(["fp16", "int8", "int4"])
+    def test_exporters_cli_eagle3_vlm_quantization(self, weight_format: str):
+        # Regression test: exporting the VLM-flavored Eagle3 draft model (e.g. AngelSlim's
+        # Qwen3-VL eagle3 checkpoints) with any weight format used to fail with
+        # `KeyError: 'input_ids'` in `generate_dummy_inputs` because `eagle3_vlm` configs
+        # replace `input_ids` with `inputs_embeds` (see model_configs.py LlamaOpenVINOConfig).
+        model_type = "qwen3_vl_eagle3"
+        task = "text-generation-with-past"
+        with TemporaryDirectory() as tmpdir:
+            add_ops = "--group-size 16" if weight_format == "int4" else ""
+            subprocess.run(
+                f"optimum-cli export openvino --model {MODEL_NAMES[model_type]} --task {task} "
+                f"--trust-remote-code --weight-format {weight_format} {add_ops} {tmpdir}",
+                shell=True,
+                check=True,
+            )
+            # Must be loaded with OVModelForCausalLM: it is a standalone draft causal LM,
+            # not a multi-component VLM, even though its config carries VLM-oriented fields.
+            model = OVModelForCausalLM.from_pretrained(tmpdir, use_cache=True)
+            self.assertTrue(model.stateful)
+
+            # Loading the same export with OVModelForVisualCausalLM must raise a clear,
+            # actionable error instead of a bare `KeyError: 'llama'`.
+            with self.assertRaisesRegex(ValueError, "OVModelForCausalLM"):
+                OVModelForVisualCausalLM.from_pretrained(tmpdir, use_cache=True)
+
     @parameterized.expand(SUPPORTED_SD_HYBRID_ARCHITECTURES)
     def test_exporters_cli_hybrid_quantization(
         self, model_type: str, expected_fake_nodes: int, expected_int8_nodes: int