diff --git a/benchmarks/single_node/fixed_seq_len/qwen3.827b_bf16_mi300x_mtp.sh b/benchmarks/single_node/fixed_seq_len/qwen3.827b_bf16_mi300x_mtp.sh
new file mode 100755
index 0000000000..c98da1b09d
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/qwen3.827b_bf16_mi300x_mtp.sh
@@ -0,0 +1,132 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+# Qwen3.8-27B (native BF16 checkpoint of the dense hybrid
+# attention model: 48 linear-attention and 16 full attention layers) on one
+# MI300X, served by vLLM with the checkpoint's own MTP head (one MTP layer,
+# mtp_num_hidden_layers=1) drafting three tokens per step.
+# https://recipes.vllm.ai/Qwen/Qwen3.8-27B
+# https://huggingface.co/Qwen/Qwen3.8-27B
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC ISL OSL RANDOM_RANGE_RATIO RESULT_FILENAME MAX_MODEL_LEN EVAL_ONLY RUN_EVAL THINKING_MODE
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+  check_env_vars SLURMD_NODENAME
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$TP" -ne 1 ]]; then
+  echo "This recipe serves Qwen3.8-27B on a single GPU; got TP=$TP" >&2
+  exit 1
+fi
+
+# The measured BF16 AL below belongs only to thinking-on, three-token MTP.
+if [[ "$THINKING_MODE" != thinking_on ]]; then
+    echo "This recipe requires THINKING_MODE=thinking_on; got $THINKING_MODE" >&2
+    exit 1
+fi
+
+# Keep the source FP8 recipe's three-token native MTP workload with BF16 weights.
+NUM_SPEC_TOKENS=3
+
+# gfx942: the ROCm vLLM arms run Triton attention and keep the KV cache in bf16
+# (no calibrated fp8 attention scales on this card); breakable piecewise graph
+# capture is off as in the MiniMax-M3 and DSv4.1 Flash gfx942 arms.
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
+export PYTHONNOUSERSITE=1
+
+rocm-smi --showmeminfo vram || true
+
+# Complete/resume partial downloads instead of trusting nonempty directories.
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+SERVER_LOG=/workspace/server.log
+
+# Serve the matrix context (isl + osl + slack), not the checkpoint's 262K;
+# accuracy evals use the eval context benchmark_lib derives.
+MODEL_LEN="$MAX_MODEL_LEN"
+if [[ "$EVAL_ONLY" == true ]]; then
+    setup_eval_context
+    MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+# vLLM's default max-num-seqs (1024) exceeds the GDN/Mamba cache blocks that fit
+# next to the weights on the smaller cards (472 next to the bf16 weights on an
+# 80 GB H100, run 35357364404) and engine start aborts before graph capture.
+# Size the scheduler batch to the sweep point instead; the accuracy eval serves
+# up to 256 concurrent requests.
+MAX_NUM_SEQS=$(( CONC > 16 ? CONC : 16 ))
+if [[ "$EVAL_ONLY" == true ]]; then
+    MAX_NUM_SEQS=256
+fi
+
+# Pyxis shares the host network; port 8888 can already belong to a host service.
+select_available_server_port
+
+# Native MTP: no draft model, the head ships in the checkpoint (mtp.* tensors).
+# BF16 thinking_on[3] from the measured SPEED-Bench curve in PR #3304.
+# Any accuracy run keeps real verification, including combined throughput/eval.
+if [[ "$EVAL_ONLY" == true || "$RUN_EVAL" == true ]]; then
+    SPEC_CONFIG=$(printf '{"method":"mtp","num_speculative_tokens":%d}' "$NUM_SPEC_TOKENS")
+else
+    SPEC_CONFIG=$(printf '{"method":"mtp","num_speculative_tokens":%d,"rejection_sample_method":"synthetic","synthetic_acceptance_length":2.51}' "$NUM_SPEC_TOKENS")
+fi
+
+start_gpu_monitor
+
+VLLM_CMD=(
+    vllm serve "$MODEL" --served-model-name "$MODEL"
+    --host 0.0.0.0 --port "$PORT"
+    --tensor-parallel-size 1
+    # Text-only serving: skip the vision tower of Qwen3_5ForConditionalGeneration.
+    --language-model-only
+    --trust-remote-code
+    # Target and native MTP head retain BF16 weights, compute and KV cache.
+    --dtype bfloat16
+    --kv-cache-dtype auto
+    --attention-backend TRITON_ATTN
+    --max-model-len "$MODEL_LEN"
+    --max-num-seqs "$MAX_NUM_SEQS"
+    # Every fixed-sequence request prefills its full prompt; no prefix-cache hits.
+    --no-enable-prefix-caching
+    --reasoning-parser qwen3
+    --default-chat-template-kwargs '{"enable_thinking":true}'
+    --enable-auto-tool-choice --tool-call-parser qwen3_xml
+    --speculative-config "$SPEC_CONFIG"
+    --disable-uvicorn-access-log
+)
+printf '%q ' "${VLLM_CMD[@]}" | tee /workspace/vllm_command.txt
+printf '\n' | tee -a /workspace/vllm_command.txt
+"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [[ "$EVAL_ONLY" == true ]]; then
+    run_eval --framework lm-eval --port "$PORT"
+    # Non-agentic evals must stage lm-eval's output into the workspace root.
+    append_lm_eval_summary
+else
+    pip install -q datasets pandas
+    # The checkpoint chat template defaults to thinking on for client-side prompts.
+    run_benchmark_serving \
+        --model "$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 "$RESULT_FILENAME" \
+        --result-dir /workspace/ \
+        --use-chat-template \
+        --server-pid "$SERVER_PID"
+    if [[ "$RUN_EVAL" == true ]]; then
+        run_eval --framework lm-eval --port "$PORT"
+        append_lm_eval_summary
+    fi
+fi
+
+stop_gpu_monitor
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 9b4fcadeb7..1ebde95c52 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1472,3 +1472,22 @@ dsv41flash-fp4-mi355x-vllm-agentic-dspark:
       search-space:
       # Follow upstream AMD Engram defaults; omit the CUDA-only config flag.
       - { tp: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32] }
+
+qwen3.827b-bf16-mi300x-vllm-mtp:
+  image: vllm/vllm-openai-rocm:nightly-a8d1aa9c99b8698a2a78b611b7a10c30e6b3995b
+  model: Qwen/Qwen3.8-27B
+  model-prefix: qwen3.827b
+  runner: cluster:mi300x-amd
+  precision: bf16
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 1, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 1, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
diff --git a/docs/configuration-procedures.md b/docs/configuration-procedures.md
index 4c6abfccfa..701fc99b07 100644
--- a/docs/configuration-procedures.md
+++ b/docs/configuration-procedures.md
@@ -222,6 +222,19 @@ Sources: [`AGENTS.md#non-negotiable-benchmark-invariants`](../AGENTS.md#non-nego
 7. Add script + master entry + launcher routing + changelog together.
 8. Run Bash syntax and generation checks. Inspect `spec-decoding`, draft/native method, token count, chat-template use, capture range, and resolved script.
 
+### Qwen3.8-27B BF16 native MTP
+
+The MI300X recipe supports 1k/1k and 8k/1k at TP1, with concurrency 1–128.
+
+The BF16 vLLM recipes use thinking on with three native MTP draft tokens. Throughput
+uses `rejection_sample_method: synthetic` and `synthetic_acceptance_length: 2.51`,
+from the BF16 `thinking_on[3]` measurement in [#3304](https://github.com/SemiAnalysisAI/InferenceX/pull/3304).
+The server explicitly enables thinking in its default chat-template kwargs; the
+fixed-sequence client uses the checkpoint's thinking-on default via `--use-chat-template`.
+The recipes require `THINKING_MODE=thinking_on` so another mode cannot reuse this AL.
+Runs requesting accuracy through either `EVAL_ONLY` or `RUN_EVAL` use real MTP
+verification. Target weights, native MTP weights and KV cache retain BF16.
+
 ### DeepSeek-V4.1-Flash DSpark
 
 The GB200 DSpark recipe uses a minimum CUDA graph capture size of 64 tokens to cover concurrent AgentX subagents. This raises c1/c2/c4 from 8/16/32 to 64; c8 and above retain their existing sizes. The full trace, AL 3.51, and Engram UVA settings are preserved; low-concurrency tail latency improvements require CI confirmation.
diff --git a/docs/configuration-procedures_zh.md b/docs/configuration-procedures_zh.md
index f877035f18..6b00d29d55 100644
--- a/docs/configuration-procedures_zh.md
+++ b/docs/configuration-procedures_zh.md
@@ -199,6 +199,19 @@ llm-d 不是 srt-slurm 路径：InferenceX 自己持有 Slurm allocation，并
 7. 同时添加脚本 + 主配置条目 + launcher 路由 + changelog。
 8. 运行 Bash 语法和生成检查；检查 `spec-decoding`、draft/native 方法、token 数、chat-template 使用、capture 范围和解析出的脚本。
 
+### Qwen3.8-27B BF16 原生 MTP
+
+MI300X 配方支持 TP1 的 1k/1k 和 8k/1k，并发范围为 1–128。
+
+BF16 vLLM 配方启用 thinking，使用 3 个原生 MTP 草稿 token。吞吐测试设置
+`rejection_sample_method: synthetic` 和 `synthetic_acceptance_length: 2.51`，
+对应 [#3304](https://github.com/SemiAnalysisAI/InferenceX/pull/3304) 中 BF16 的 `thinking_on[3]` 测量值。
+服务端在默认 chat-template kwargs 中显式开启 thinking；固定序列长度客户端通过
+`--use-chat-template` 使用 checkpoint 默认开启的 thinking 模式。
+配方要求 `THINKING_MODE=thinking_on`，避免其他模式误用此 AL。
+通过 `EVAL_ONLY` 或 `RUN_EVAL` 请求准确率评测时，使用真实 MTP 验证。
+目标权重、原生 MTP 权重及 KV cache 均保持 BF16。
+
 ### DeepSeek-V4.1-Flash DSpark
 
 GB200 的 DSpark 配方将 CUDA graph 最小捕获范围设为 64 tokens，以覆盖 AgentX 子代理并发。这会将 c1/c2/c4 的上限从 8/16/32 提升至 64；c8 及以上保持原有大小。完整轨迹、AL 3.51 和 Engram UVA 配置保持不变；需通过 CI 验证低并发尾延迟改善。
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index c5b4309f5f..b85b2619d6 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8455,3 +8455,24 @@
   description:
     - "Update B200 vLLM AgentX to DSpark6 and a new image with TP8 and DEP8 configurations."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3274
+
+- config-keys:
+    - qwen3.827b-bf16-mi300x-vllm-mtp
+  description:
+    - "Add MI300X Qwen3.8-27B BF16 vLLM TP1 native MTP with three draft tokens, BF16 KV cache, chat-templated 1k1k prompts, and concurrency 1-128; rerun the workload from #3294 with the unquantized checkpoint and the same pinned image"
+    - "新增 MI300X Qwen3.8-27B BF16 vLLM TP1 原生 MTP 配方：预测 3 个 token，使用 BF16 KV cache、经 chat template 处理的 1k1k 提示词及并发 1-128；使用原始未量化 checkpoint 和相同固定镜像重跑 #3294 的工作负载"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3307
+
+- config-keys:
+    - qwen3.827b-bf16-mi300x-vllm-mtp
+  description:
+    - "Use thinking-on synthetic AL 2.51 with three native MTP draft tokens for MI300X Qwen3.8-27B BF16 throughput, matching the BF16 SPEED-Bench curve in #3304; accuracy runs retain real verification"
+    - "MI300X Qwen3.8-27B BF16 吞吐测试使用 thinking 开启、3 个原生 MTP 草稿 token 和合成 AL 2.51，对应 #3304 的 BF16 SPEED-Bench 曲线；准确率评测保留真实验证"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3307
+
+- config-keys:
+    - qwen3.827b-bf16-mi300x-vllm-mtp
+  description:
+    - "Add 8k1k at concurrency 1-128 for MI300X Qwen3.8-27B BF16 vLLM TP1 native MTP, retaining thinking-on synthetic AL 2.51 with three draft tokens and the existing image"
+    - "为 MI300X Qwen3.8-27B BF16 vLLM TP1 原生 MTP 新增并发 1-128 的 8k1k 场景，保持 thinking 开启、合成 AL 2.51、3 个草稿 token 及原有镜像"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3307