diff --git a/python/cactus/convert/cactus_adapters/tensor_io.py b/python/cactus/convert/cactus_adapters/tensor_io.py
index dc59d0aa8..8bfc632f1 100644
--- a/python/cactus/convert/cactus_adapters/tensor_io.py
+++ b/python/cactus/convert/cactus_adapters/tensor_io.py
@@ -292,7 +292,7 @@ def save_tensor_with_header(tensor, output_path, precision='INT8', transpose=Fal
             data = data * GEMMA4_WEIGHT_SCALE
         elif 'router_scale' in filename:
             data = data / GEMMA4_WEIGHT_SCALE
-        elif filename in ('token_embeddings.weights', 'output_weight.weights', 'embed_vision_embedding.weights', 'embed_vision_proj.weights'):
+        elif filename in ('token_embeddings.weights', 'output_weight.weights', 'embed_vision_proj.weights', 'embed_audio_proj.weights'):
             data = data / GEMMA4_WEIGHT_SCALE
         elif filename == 'output_norm.weights':
             data = data * GEMMA4_WEIGHT_SCALE
diff --git a/python/cactus/convert/model_adapters/naming.py b/python/cactus/convert/model_adapters/naming.py
index d854b1a02..e96f568fb 100644
--- a/python/cactus/convert/model_adapters/naming.py
+++ b/python/cactus/convert/model_adapters/naming.py
@@ -23,7 +23,6 @@ class NameMatch:
 GEMMA4_DIV_BASENAMES = {
     "token_embeddings",
     "output_weight",
-    "embed_vision_embedding",
 }
 
 
diff --git a/python/cactus/transpile/model_adapters.py b/python/cactus/transpile/model_adapters.py
index b09357e42..b6e2cee0d 100644
--- a/python/cactus/transpile/model_adapters.py
+++ b/python/cactus/transpile/model_adapters.py
@@ -1094,10 +1094,6 @@ def _gemma4_compute_native_like_audio_features(
     if not callable(getattr(embed_audio, "forward", None)):
         raise TypeError("Gemma4 audio embedder is not callable")
     projected = embed_audio(inputs_embeds=audio_encodings)
-    # Native Cactus Gemma4 scales audio soft tokens before merging them into
-    # the language stream. HF leaves this implicit in its fp16 path, but the
-    # quantized Cactus decoder expects the native-scaled representation.
-    projected = projected * (1.0 / 16.0)
     # The transpiled bundle is shape-specialized from representative media.
     # Native Gemma4 audio preprocessing emits an unpadded feature tensor for
     # that media, so the post-subsampling sequence is already the real token