diff --git a/x/create/gemma4.go b/x/create/gemma4.go
index 35e920077..2bb493db3 100644
--- a/x/create/gemma4.go
+++ b/x/create/gemma4.go
@@ -93,6 +93,13 @@ func (t gemma4ImportTransform) quantizationType(name string, shape []int32, quan
 		return ""
 	}
 
+	// MoE router logits choose the top-k expert set. Quantization noise here
+	// can flip expert selection, after which downstream activations diverge
+	// sharply. The tensor is small, so leave it in source precision.
+	if isGemma4RouterProjection(name) {
+		return ""
+	}
+
 	// Mixed-precision quantization: sensitive tensors get higher precision.
 	//
 	// Value projections (v_proj) directly determine attention output quality.
@@ -170,6 +177,12 @@ func isEmbedTokensWeight(name string) bool {
 		!strings.Contains(name, "per_layer")
 }
 
+func isGemma4RouterProjection(name string) bool {
+	return strings.HasSuffix(name, ".router.proj.weight") &&
+		!strings.Contains(name, "audio_tower") &&
+		!strings.Contains(name, "vision_tower")
+}
+
 func (t gemma4ImportTransform) transformTensor(td *safetensors.TensorData) ([]*safetensors.TensorData, error) {
 	if td == nil {
 		return nil, nil
diff --git a/x/create/gemma4_test.go b/x/create/gemma4_test.go
index 40b183162..858183db0 100644
--- a/x/create/gemma4_test.go
+++ b/x/create/gemma4_test.go
@@ -68,6 +68,11 @@ func TestGemma4QuantizationType(t *testing.T) {
 		{"expert gate_up nvfp4", transform26B, "model.layers.0.moe.experts.42.gate_up_proj.weight", aligned, "nvfp4", "nvfp4"},
 		{"expert gate_up mxfp4", transform26B, "model.layers.0.moe.experts.42.gate_up_proj.weight", aligned, "mxfp4", "mxfp4"},
 
+		// === Router projection: expert selection is sensitive; keep source precision ===
+		{"router proj int4", transform26B, "model.layers.0.router.proj.weight", aligned, "int4", ""},
+		{"router proj nvfp4", transform26B, "model.layers.0.router.proj.weight", aligned, "nvfp4", ""},
+		{"router proj mxfp4", transform26B, "model.layers.0.router.proj.weight", aligned, "mxfp4", ""},
+
 		// === k_proj: promoted only for 8-expert models ===
 		{"k_proj 128 experts int4", transform26B, "model.layers.0.self_attn.k_proj.weight", aligned, "int4", "int4"},
 		{"k_proj 8 experts int4", transform8E, "model.layers.0.self_attn.k_proj.weight", aligned, "int4", "int8"},