diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_vllm.sh
index 011f9dedb..718b11b8b 100755
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b200_vllm.sh
@@ -37,16 +37,16 @@ if [ "${DP_ATTENTION}" = "true" ]; then
 fi
 
 EP_ARGS=()
+MOE_ARGS=()
 if [ "${EP_SIZE:-1}" -gt 1 ]; then
     EP_ARGS=(--enable-expert-parallel)
+    MOE_ARGS=(--moe-backend deep_gemm_mega_moe)
 fi
 
 GMU_ARGS=()
-MOE_ARGS=()
 EPLB_ARGS=()
 PREFILL_SCHEDULE_ARGS=()
 if [ "${DP_ATTENTION}" = "true" ]; then
-    MOE_ARGS=(--moe-backend deep_gemm_mega_moe)
     EPLB_ARGS=(--enable-eplb --eplb-config '{"communicator":"torch_nccl", "use_async": false}')
     PREFILL_SCHEDULE_ARGS=(--prefill-schedule-interval 4)
 fi
@@ -60,6 +60,10 @@ fi
 MAX_CUDAGRAPH_CAPTURE_SIZE=2048
 
 BENCHMARK_MAX_MODEL_LEN="$MAX_MODEL_LEN"
+# Cap MAX_MODEL_LEN to free CUDA-graph memory and give more headroom for KV blocks
+if [ "${BENCHMARK_MAX_MODEL_LEN}" -gt 12288 ]; then
+    BENCHMARK_MAX_MODEL_LEN=12288
+fi
 
 if [ "${EVAL_ONLY}" = "true" ]; then
     EVAL_MAX_MODEL_LEN=$(compute_eval_context_length "$MODEL" "$BENCHMARK_MAX_MODEL_LEN")
@@ -91,6 +95,7 @@ vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
     --enable-auto-tool-choice \
     --reasoning-parser deepseek_v4 \
     --max-cudagraph-capture-size "$MAX_CUDAGRAPH_CAPTURE_SIZE" \
+    --gpu-memory-utilization 0.95 \
     --max-model-len "$SERVE_MAX_MODEL_LEN" \
     --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" > "$SERVER_LOG" 2>&1 &
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index c752b04f6..661a05b99 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -929,7 +929,23 @@ dsv4-fp4-b200-sglang:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 1024 }
 
 dsv4-fp4-b200-vllm:
-  image: vllm/vllm-openai:v0.25.0
+  image: vllm/vllm-openai:nightly-821717118fc26667dd474b9b0ab81d29259dfc5c
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:b200-dgxc
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, conc-start: 1, conc-end: 256 }
+      - { tp: 8, ep: 8, conc-start: 1, conc-end: 256 }
+      - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 1024 }
+m:
+  image: vllm/vllm-openai:v0.26.0
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:b200-dgxc
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0bce8d894..e54ceaf39 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5818,3 +5818,9 @@
   description:
     - "Extend the SimpleCPUOffloadConnector grid to c8/c12/c16/c20/c24/c28/c32/c48/c64 to locate its crossover against the resident curve"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2475
+
+- config-keys:
+    - dsv4-fp4-b200-vllm
+  description:
+    - "B200 DSv4 FP4 vLLM: nightly image, new sweep (TP8, TEP8, DEP8), megamoe when ep>1, gmu 0.95, cap max-model-len 12288"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2534