diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 7e975fdba..f161e9bdc 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2510,6 +2510,27 @@ dsv4-fp8-h200-vllm:
     search-space:
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64 }
 
+# MTP variant of dsv4-fp8-h200-vllm. Uses the canonical v0.20.0-cu130 image
+# (the non-MTP entry above is still on the deepseekv4-cu129 tag) and adds
+# --speculative-config '{"method":"mtp","num_speculative_tokens":2}'.
+dsv4-fp8-h200-vllm-mtp:
+  image: vllm/vllm-openai:v0.20.0-cu130
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: h200
+  precision: fp8
+  framework: vllm
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64, spec-decoding: mtp }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 4, conc-end: 64, spec-decoding: mtp }
+
 # DeepSeek-V4-Pro B300 single-node aggregate recipe from the submitted B300
 # pareto sweep. The single-node schema has no explicit data-parallel-size
 # field, so dp-attn=true is used as the existing vLLM script switch for DP4
diff --git a/benchmarks/single_node/dsv4_fp8_h200_mtp.sh b/benchmarks/single_node/dsv4_fp8_h200_mtp.sh
new file mode 100755
index 000000000..5a6834757
--- /dev/null
+++ b/benchmarks/single_node/dsv4_fp8_h200_mtp.sh
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+
+# DeepSeek-V4-Pro H200 vLLM MTP variant of the recipe at
+# https://vllm.ai/blog/deepseek-v4. Mirrors dsv4_fp8_h200.sh but adds
+# --speculative-config '{"method":"mtp","num_speculative_tokens":1}' and
+# routes prompts through chat-formatted encoding via --dsv4 (required for
+# meaningful MTP acceptance numbers per AGENTS.md).
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+hf download "$MODEL"
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+# DeepSeek-V4-Pro weights are large; engine startup can exceed the default
+# 600s. Give it an hour to load.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+# Skip the cudagraph-memory estimator during the worker memory profiling
+# phase — it overestimates and pushes us over the GPU memory budget on
+# H200 + MTP, even though the actual cudagraph capture works fine.
+export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN_ARG="--max-model-len $EVAL_MAX_MODEL_LEN"
+else
+    MAX_MODEL_LEN_ARG="--max-model-len $MAX_MODEL_LEN"
+fi
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+# Per the recipe, run with EP + DP=8 (no --tensor-parallel-size flag). TP
+# from the search space is used only for GPU allocation by the runner and
+# as the DP size.
+set -x
+vllm serve $MODEL --host 0.0.0.0 --port $PORT \
+--trust-remote-code \
+--kv-cache-dtype fp8 \
+--block-size 256 \
+--no-enable-prefix-caching \
+--enable-expert-parallel \
+--data-parallel-size $TP \
+$MAX_MODEL_LEN_ARG \
+--gpu-memory-utilization 0.95 \
+--max-num-seqs 512 \
+--max-num-batched-tokens 512 \
+--no-enable-flashinfer-autotune \
+--compilation-config '{"mode":0,"cudagraph_mode":"FULL_DECODE_ONLY"}' \
+--speculative-config '{"method":"mtp","num_speculative_tokens":1}' \
+--tokenizer-mode deepseek_v4 \
+--tool-call-parser deepseek_v4 \
+--enable-auto-tool-choice \
+--reasoning-parser deepseek_v4 > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+# MTP acceptance rate degrades on raw random tokens; --dsv4 routes prompts
+# through chat-formatted encoding as required for speculative decoding benchmarks.
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --dsv4
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2bd14e776..7009a1b8b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1985,3 +1985,13 @@
     - "Topology: 1 prefill DEP8 worker and 4 decode TP8 workers with dedicated NATS/etcd"
     - "Mirrors the historical 1P4D DEP8/TP8 offload point from srt-slurm aflowers/vllm-gb200-v0.20.0"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1218
+
+- config-keys:
+    - dsv4-fp8-h200-vllm-mtp
+  description:
+    - "Add DeepSeek-V4-Pro FP8 H200 vLLM MTP variant (mirrors dsv4-fp8-h200-vllm with --speculative-config {\"method\":\"mtp\",\"num_speculative_tokens\":1})"
+    - "Image: vllm/vllm-openai:v0.20.0-cu130"
+    - "Set VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0 to skip the cudagraph-memory estimator (it overshoots the H200 + MTP memory budget at profile time even though actual cudagraph capture works fine)"
+    - "run_benchmark_serving uses --dsv4 (chat-formatted prompts) per the AGENTS.md MTP rule, since EAGLE-style speculative decoding regresses acceptance on raw random tokens"
+    - "Search space mirrors the non-MTP H200 entry: TP=8, EP=8, DP-attn=true, CONC 4-64 for both 1k1k and 8k1k, with spec-decoding: mtp"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1222