diff --git a/python/sglang/srt/models/mimo_v2.py b/python/sglang/srt/models/mimo_v2.py
index 26f9c8c88fdf..41a95e4b6e2f 100644
--- a/python/sglang/srt/models/mimo_v2.py
+++ b/python/sglang/srt/models/mimo_v2.py
@@ -1390,16 +1390,11 @@ def forward(
         )
 
         aux_hidden_states = None
-        if self.capture_aux_hidden_states:
-            hidden_states, hidden_states_before_norm, aux_hidden_states = self.model(
-                input_ids,
-                positions,
-                forward_batch,
-                input_embeds,
-                pp_proxy_tensors=pp_proxy_tensors,
-            )
-        elif self._is_multimodal:
-            hidden_states, hidden_states_before_norm = general_mm_embed_routine(
+        # Multimodal embedding must run even when aux hidden states are captured
+        # (DFLASH/MTP). Taking the capture branch first skipped image/audio
+        # feature placement and left placeholder tokens as plain text embeddings.
+        if self._is_multimodal:
+            model_out = general_mm_embed_routine(
                 input_ids=input_ids,
                 forward_batch=forward_batch,
                 language_model=self.model,
@@ -1408,13 +1403,17 @@ def forward(
                 pp_proxy_tensors=pp_proxy_tensors,
             )
         else:
-            hidden_states, hidden_states_before_norm = self.model(
+            model_out = self.model(
                 input_ids,
                 positions,
                 forward_batch,
                 input_embeds,
                 pp_proxy_tensors=pp_proxy_tensors,
             )
+        if self.capture_aux_hidden_states:
+            hidden_states, hidden_states_before_norm, aux_hidden_states = model_out
+        else:
+            hidden_states, hidden_states_before_norm = model_out
 
         if self.pp_group.is_last_rank:
             return self.logits_processor(
diff --git a/python/sglang/srt/models/mimo_vl.py b/python/sglang/srt/models/mimo_vl.py
index 3f4c7267bc20..f09269adc33f 100644
--- a/python/sglang/srt/models/mimo_vl.py
+++ b/python/sglang/srt/models/mimo_vl.py
@@ -317,6 +317,10 @@ def __init__(
             prefix=add_prefix("merger", prefix),
             use_data_parallel=self.use_data_parallel,
         )
+        # MiMo-VL merger ln_q is LayerNorm (see modeling_mimo_v2.py
+        # MiMoVisionPatchMerger), not the Qwen2.5-VL RMSNorm. Checkpoint ships
+        # visual.merger.ln_q.weight only (bias omitted → zeros).
+        self.merger.ln_q = nn.LayerNorm(hidden_size, eps=1e-6, bias=False)
         self._post_init()
 
     def apply_index(self, tensor: torch.Tensor, index: torch.Tensor):
@@ -327,7 +331,9 @@ def apply_index(self, tensor: torch.Tensor, index: torch.Tensor):
 
     def _post_init(self):
         for name, param in self.named_parameters():
-            if "bias" in name:
+            # Also zero sinks: they are torch.empty at ctor and some checkpoints
+            # omit individual visual.blocks.*.attn.sinks keys.
+            if "bias" in name or name.endswith("sinks"):
                 param.data.zero_()
 
     def get_window_index_1d(self, grid_thw, col=True):
diff --git a/python/sglang/srt/multimodal/processors/mimo_v2.py b/python/sglang/srt/multimodal/processors/mimo_v2.py
index e914f00b1a0e..35a43469fdcf 100644
--- a/python/sglang/srt/multimodal/processors/mimo_v2.py
+++ b/python/sglang/srt/multimodal/processors/mimo_v2.py
@@ -214,8 +214,12 @@ def __post_init__(self):
                 )
 
 
-_QWEN2VL_PIXEL_MEAN = torch.Tensor([123.675, 116.28, 103.53]).view(-1, 1, 1)
-_QWEN2VL_PIXEL_STD = torch.Tensor([58.395, 57.12, 57.375]).view(-1, 1, 1)
+# OPENAI_CLIP stats, matching Qwen2VLImageProcessor (preprocessor_config.json
+# image_mean/image_std with do_rescale 1/255). Kept in 0-255 space because
+# standardize_batch sees un-rescaled PIL tensors. ImageNet values were wrong
+# for MiMo-VL and scrambled colors.
+_QWEN2VL_PIXEL_MEAN = torch.Tensor([0.48145466, 0.4578275, 0.40821073]).view(-1, 1, 1) * 255.0
+_QWEN2VL_PIXEL_STD = torch.Tensor([0.26862954, 0.26130258, 0.27577711]).view(-1, 1, 1) * 255.0
 _mean_std_cache = {}