diff --git a/clusters/k3s-dgx/nim-service/qwen.yaml b/clusters/k3s-dgx/nim-service/qwen.yaml
index 872b8e9..ef146df 100644
--- a/clusters/k3s-dgx/nim-service/qwen.yaml
+++ b/clusters/k3s-dgx/nim-service/qwen.yaml
@@ -49,7 +49,7 @@ spec:
     - --served-model-name
     - Qwen/Qwen3.6-27B-FP8
     - --gpu-memory-utilization
-    - "0.92"                  # ↑ More memory for KV cache / bigger batches
+    - "0.90"                  # ↑ More memory for KV cache / bigger batches
     - --max-model-len
     - "32768"                 # ↓ 256K→32K (adjust if you really need long context)
     - --language-model-only