diff --git a/benchmarks/single_node/agentic/qwen3.8next_fp4_b200_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.8next_fp4_b200_sglang_mtp.sh
index ef23585e29..43b792b892 100755
--- a/benchmarks/single_node/agentic/qwen3.8next_fp4_b200_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.8next_fp4_b200_sglang_mtp.sh
@@ -156,7 +156,9 @@ SGLANG_CMD=(
     # NEXTN silently resets --max-running-requests to 48 when it is unset, so
     # this must stay explicit and sized to the AgentX concurrency.
     --max-running-requests "$MAX_RUNNING_REQUESTS"
-    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    # SGLang v0.5.20 removed the deprecated --cuda-graph-max-bs alias
+    # (sgl-project/sglang#38375); the decode-phase flag is its replacement.
+    --cuda-graph-max-bs-decode "$CUDA_GRAPH_MAX_BS"
     --mem-fraction-static "$MEM_FRACTION_STATIC"
     "${MAMBA_CACHE_ARGS[@]}"
     --stream-interval 50
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index c5d5edac68..a36adc0533 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -5198,7 +5198,7 @@ qwen3.5-fp4-b200-sglang-agentic-mtp:
 # Qwen3.8-Flash-Next NVFP4 AgentX on B200 via SGLang with native NEXTN MTP.
 # TP1 matches the verified single-node recipe.
 qwen3.8next-fp4-b200-sglang-agentic-mtp:
-  image: lmsysorg/sglang:qwen38flashnext
+  image: lmsysorg/sglang:v0.5.20-cu130
   model: RadixArk/Qwen3.8-Flash-Next-NVFP4
   model-prefix: qwen3.8next
   runner: cluster:b200-nscale