diff --git a/tensorrt_llm/_torch/models/_arch_index.py b/tensorrt_llm/_torch/models/_arch_index.py
index f2f278dea788..d6a98a4908ca 100644
--- a/tensorrt_llm/_torch/models/_arch_index.py
+++ b/tensorrt_llm/_torch/models/_arch_index.py
@@ -89,6 +89,7 @@ def is_builtin_zoo_module(module_name: str) -> bool:
     "NemotronHPuzzleForCausalLM": "modeling_nemotron_h",
     "NemotronH_Nano_Omni_Reasoning_V3": "modeling_nemotron_nano",
     "NemotronH_Nano_VL_V2": "modeling_nemotron_nano",
+    "NemotronH_Omni_Reasoning_V3": "modeling_nemotron_nano",
     "Phi3ForCausalLM": "modeling_phi3",
     "Phi4MMForCausalLM": "modeling_phi4mm",
     "PixtralForConditionalGeneration": "modeling_mistral",
@@ -215,6 +216,7 @@ def is_builtin_zoo_module(module_name: str) -> bool:
     "mistral3": "modeling_mistral",
     "mistral_common": "modeling_mistral",
     "mistral_large_3": "modeling_mistral",
+    "nemotron_h_omni": "modeling_nemotron_nano",
     "phi4mm": "modeling_phi4mm",
     "qwen2_5_vl": "modeling_qwen2vl",
     "qwen2_vl": "modeling_qwen2vl",
diff --git a/tensorrt_llm/_torch/models/modeling_nemotron_nano.py b/tensorrt_llm/_torch/models/modeling_nemotron_nano.py
index ed0808b30d2e..ac0ea865e75e 100644
--- a/tensorrt_llm/_torch/models/modeling_nemotron_nano.py
+++ b/tensorrt_llm/_torch/models/modeling_nemotron_nano.py
@@ -1029,8 +1029,10 @@ def __init__(
         self.img_context_token = self.config.img_context_token
         self.video_context_token = self.config.video_context_token
         self.video_context_token_id = self.config.video_context_token_id
-        self.img_start_token = self.config.img_start_token
-        self.img_end_token = self.config.img_end_token
+        # Nemotron 3.5 Super VL carries only `img_context_token`; fall back to the
+        # InternVL `<img>`/`</img>` pair its processor uses.
+        self.img_start_token = getattr(self.config, "img_start_token", "<img>")
+        self.img_end_token = getattr(self.config, "img_end_token", "</img>")
         # Pre-tokenize special tokens for video EVS processing (following vLLM).
         # These may be multi-token under BPE, so we store the full ID list.
         self._img_start_token_ids = self.tokenizer.encode(
@@ -1068,8 +1070,13 @@ def __init__(
                     "Dynamic resolution (enabled via `vision_config.min_num_patches`) only supports "
                     f"`config.ps_version='v2'. Got {pixel_shuffle_version=}."
                 )
+            # Nemotron 3.5 Super VL states the context length only on the inner
+            # `llm_config`, with no top-level `max_sequence_length`.
+            max_model_len = getattr(config, "max_sequence_length", None)
+            if max_model_len is None:
+                max_model_len = config.llm_config.max_position_embeddings
             self.dynamic_tiler = DynamicResolutionImageTiler(
-                max_model_len=config.max_sequence_length,
+                max_model_len=max_model_len,
                 patch_size=self.patch_size,
                 downsample_ratio=self.downsample_ratio,
                 min_num_patches=vision_args["min_num_patches"],
@@ -1082,12 +1089,17 @@ def __init__(
         # Video temporal compression and sizing config.
         vision_config = getattr(config, "vision_config", config)
         self.video_temporal_patch_size = getattr(vision_config, "video_temporal_patch_size", 1)
+        # Tile-based (InternVL-style) checkpoints state the video sizing on
+        # `vision_config`. Patch-based ones (Nemotron 3.5 Super VL) state it on the
+        # image processor instead, and that processor has no `max_num_tiles`, so the
+        # tile fallback in `_process_videos_frames` cannot serve them.
+        video_sizing = vision_config if hasattr(self.processor, "max_num_tiles") else self.processor
         self.video_maintain_aspect_ratio = getattr(
-            vision_config, "video_maintain_aspect_ratio", False
+            video_sizing, "video_maintain_aspect_ratio", False
         )
         # Resolve video target size: video_target_num_patches or video_target_img_size.
-        target_num_patches = getattr(vision_config, "video_target_num_patches", None)
-        target_img_size = getattr(vision_config, "video_target_img_size", None)
+        target_num_patches = getattr(video_sizing, "video_target_num_patches", None)
+        target_img_size = getattr(video_sizing, "video_target_img_size", None)
         if target_num_patches is not None and target_img_size is not None:
             raise ValueError(
                 "Exactly one of video_target_num_patches or "
@@ -2613,6 +2625,9 @@ def _resample_audios(
 @register_vision_encoder(NanoV2VLMultimodalEncoder)
 @register_auto_model("NemotronH_Nano_Omni_Reasoning_V3")
 @register_auto_model("NemotronH_Nano_VL_V2")
+# Nemotron 3.5 Super VL: same class, no "Nano" infix, and vision-only (its
+# `sound_config` is null) despite the "Omni" in the architecture name.
+@register_auto_model("NemotronH_Omni_Reasoning_V3")
 @register_input_processor(
     NanoV2VLInputProcessor,
     model_type="NemotronH_Nano_VL_V2",
@@ -2623,6 +2638,13 @@ def _resample_audios(
     model_type="NemotronH_Nano_Omni_Reasoning_V3",
     placeholder_metadata=_NANO_VL_PLACEHOLDER_METADATA,
 )
+# `model_type`, not architecture: the Nano checkpoints set both to the same
+# string, this one does not.
+@register_input_processor(
+    NanoV2VLInputProcessor,
+    model_type="nemotron_h_omni",
+    placeholder_metadata=_NANO_VL_PLACEHOLDER_METADATA,
+)
 class NemotronH_Nano_VL_V2(MultimodalModelMixin, transformers.PreTrainedModel):
     _supports_flash_attn = True
 
diff --git a/tensorrt_llm/_torch/models/modeling_radio.py b/tensorrt_llm/_torch/models/modeling_radio.py
index 4bccf16a7841..724854bf33aa 100644
--- a/tensorrt_llm/_torch/models/modeling_radio.py
+++ b/tensorrt_llm/_torch/models/modeling_radio.py
@@ -1214,11 +1214,12 @@ def __init__(self,
         img_size = VIT_TIMM_CONFIG_BY_NAME[model_name].img_size
         mlp_ratio = intermediate_size / embed_dim
 
-        # Build the model.
+        # Build the model. C-RADIOv4-H omits `in_chans`, `input_size` and `drop`
+        # from the RADIO argument bag, so these reads tolerate their absence.
         in_chans = 3
-        if args.in_chans is not None:
+        if getattr(args, 'in_chans', None) is not None:
             in_chans = args.in_chans
-        elif args.input_size is not None:
+        elif getattr(args, 'input_size', None) is not None:
             in_chans = args.input_size[0]
         vit_model = VisionTransformer(
             img_size=img_size,
@@ -1228,7 +1229,7 @@ def __init__(self,
             depth=depth,
             num_heads=num_attention_heads,
             mlp_ratio=mlp_ratio,
-            drop_rate=args.drop,
+            drop_rate=getattr(args, 'drop', 0.0),
             special_args=args,
             model_config=self.model_config,
         )
@@ -1242,30 +1243,41 @@ def __init__(self,
         input_conditioner = nn.Identity()
         input_conditioner.dtype = config.torch_dtype
 
-        adaptor_names = config.adaptor_names or []
+        # C-RADIOv4 configs omit these three fields; absent means "not configured",
+        # which is the only case these guards accept anyway.
+        adaptor_names = getattr(config, 'adaptor_names', None) or []
         if len(adaptor_names) > 0:
             raise ValueError(
                 "Adaptor names are not supported for RADIO models.")
         adaptors = dict()
 
         feature_normalizer = None
-        if config.feature_normalizer_config is not None:
+        if getattr(config, 'feature_normalizer_config', None) is not None:
             raise ValueError(
                 "Feature normalizer is not supported for RADIO models.")
 
         inter_feature_normalizer = None
-        if config.inter_feature_normalizer_config is not None:
+        if getattr(config, 'inter_feature_normalizer_config', None) is not None:
             raise ValueError(
                 "Intermediate feature normalizer is not supported for RADIO models."
             )
 
+        # C-RADIOv4 renames both: `max_img_size` is `max_resolution`, and
+        # `image_size` the preferred (square) resolution. `vitdet_window_size`
+        # has no v4 counterpart; None is the constructor's own default.
+        max_resolution = (config.max_resolution if hasattr(
+            config, 'max_resolution') else config.max_img_size)
+        preferred_resolution = (config.preferred_resolution if hasattr(
+            config, 'preferred_resolution') else Resolution(
+                config.image_size, config.image_size))
+
         self.radio_model = RADIOVisionModelBase(
             vit_model,
             input_conditioner,
             patch_size=config.patch_size,
-            max_resolution=config.max_resolution,
-            window_size=config.vitdet_window_size,
-            preferred_resolution=config.preferred_resolution,
+            max_resolution=max_resolution,
+            window_size=getattr(config, 'vitdet_window_size', None),
+            preferred_resolution=preferred_resolution,
             adaptors=adaptors,
             feature_normalizer=feature_normalizer,
             inter_feature_normalizer=inter_feature_normalizer,
diff --git a/tests/unittest/_torch/modeling/test_modeling_nemotron_nano_v2_vl.py b/tests/unittest/_torch/modeling/test_modeling_nemotron_nano_v2_vl.py
index 993a52f3fdcb..6cc708b4b946 100644
--- a/tests/unittest/_torch/modeling/test_modeling_nemotron_nano_v2_vl.py
+++ b/tests/unittest/_torch/modeling/test_modeling_nemotron_nano_v2_vl.py
@@ -66,8 +66,12 @@ def _make_minimal_nano_model_config():
 
 @pytest.mark.cpu_only
 def test_nemotron_nano_registers_native_multimodal_epd_components():
-    """Native Nano VL/Omni classes advertise MM EPD support."""
-    for arch in ("NemotronH_Nano_VL_V2", "NemotronH_Nano_Omni_Reasoning_V3"):
+    """Every arch served by the native Nano VL class advertises MM EPD support."""
+    for arch in (
+        "NemotronH_Nano_VL_V2",
+        "NemotronH_Nano_Omni_Reasoning_V3",
+        "NemotronH_Omni_Reasoning_V3",
+    ):
         vision_encoder_cls, vlm_base_model = MODEL_CLASS_VISION_ENCODER_MAPPING[arch]
         assert vision_encoder_cls is NanoV2VLMultimodalEncoder
         assert vlm_base_model is None