diff --git a/scripts/serve-ling3-vllm.sh b/scripts/serve-ling3-vllm.sh
index d0d6084..09643d3 100755
--- a/scripts/serve-ling3-vllm.sh
+++ b/scripts/serve-ling3-vllm.sh
@@ -23,6 +23,10 @@ set +a
 export HOST="${HOST:-0.0.0.0}"
 export PORT="${PORT:-7770}"
 export SERVED_MODEL_NAME="${SERVED_MODEL_NAME:-frank-default-large}"
+# Raise the vLLM frame-buffer utilization so weights + 256k KV fit in VRAM
+# (the default 0.72 from native-256k.env OOMs; native-256k uses
+# ${GPU_MEMORY_UTILIZATION:-0.72}, so this exported value wins).
+export GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.95}"
 
 # Run from a directory with no top-level 'vllm/' dir.
 # Neither launcher changes directory, so the process cwd becomes sys.path[0]