diff --git a/scripts/vllm-awq-qwen36-dflash-ampere.sh b/scripts/vllm-awq-qwen36-dflash-ampere.sh
index 79d87c3..9835a7a 100644
--- a/scripts/vllm-awq-qwen36-dflash-ampere.sh
+++ b/scripts/vllm-awq-qwen36-dflash-ampere.sh
@@ -25,7 +25,7 @@ exec docker run --rm --name vllm-awq-qwen36-dflash-ampere \
     vllm/vllm-openai:latest "${MODEL}" \
     --served-model-name "frank-default-small" \
     --max-num-seqs 1 \
-    --max-model-len 212992 \
+    --max-model-len 32768 \
     --tensor-parallel-size 1 \
     --enable-chunked-prefill \
     --max-num-batched-tokens 4096 \