diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index bdbfafc22..f8bc53075 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2701,7 +2701,7 @@ minimaxm3-fp4-mi355x-vllm-mtp:
 # https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md
 # block size 128 is mandatory for MSA. TP4 on a single gfx950 node, per the recipe.
 minimaxm3-fp4-mi355x-atom:
-  image: rocm/atom-dev:M3
+  image: rocm/atom-dev:MiniMax-M3-20260623
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x
@@ -2713,14 +2713,30 @@ minimaxm3-fp4-mi355x-atom:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 2, conc-start: 128, conc-end: 256 }
       - { tp: 4, conc-start: 1, conc-end: 256 }
     - isl: 8192
       osl: 1024
       search-space:
-      - { tp: 2, conc-start: 128, conc-end: 256 }
       - { tp: 4, conc-start: 1, conc-end: 256 }
-      - { tp: 8, conc-start: 1, conc-end: 2 }
+
+minimaxm3-fp4-mi355x-atom-mtp:
+  image: rocm/atom-dev:MiniMax-M3-20260623
+  model: amd/MiniMax-M3-MXFP4
+  model-prefix: minimaxm3
+  runner: mi355x
+  precision: fp4
+  framework: atom
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 4, conc-start: 1, conc-end: 256, spec-decoding: mtp  }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 4, conc-start: 1, conc-end: 256, spec-decoding: mtp  }
 
 minimaxm3-fp8-mi355x-atom:
   image: rocm/atom-dev:MiniMax-M3-20260623
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
index 80f230f28..a0907f18c 100644
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
@@ -11,8 +11,7 @@ check_env_vars \
     RANDOM_RANGE_RATIO \
     RESULT_FILENAME \
     EP_SIZE \
-    DP_ATTENTION \
-    MAX_MODEL_LEN
+    DP_ATTENTION 
 
 if [[ -n "$SLURM_JOB_ID" ]]; then
   echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
@@ -22,41 +21,41 @@ echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTIO
 
 SERVER_LOG=/workspace/server.log
 
-export OMP_NUM_THREADS=1
+PARALLEL_ARGS=(-tp "$TP") #TP
+if [ "$DP_ATTENTION" = "true" ]; then
+    if [ "$EP_SIZE" -gt 1 ]; then #DP+EP
+        PARALLEL_ARGS=(-tp "$TP" --enable-expert-parallel --enable-dp-attention )
+    else #DP+TP
+        PARALLEL_ARGS=(-tp "$TP" --enable-dp-attention )
+    fi
+fi 
 
-# Use the matrix-supplied MAX_MODEL_LEN (isl + osl + 256). Eval-only jobs need a
-# larger window for the eval prompts, so override it from the eval context.
-if [ "${EVAL_ONLY}" = "true" ]; then
-    setup_eval_context
-    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
-fi
-
-if [ "$EP_SIZE" -gt 1 ]; then
-  EP=" --enable-expert-parallel"
-else
-  EP=" "
-fi
+SPEC_ARGS=()
 
 # Start GPU monitoring (power, temperature, clocks every second)
 start_gpu_monitor
 MEM_FRAC_STATIC=0.8
 
 set -x
-
-# Flags follow the ATOM MiniMax-M3 MXFP4 recipe (FP4 on 4xMI355 section):
-# https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md
-# --block-size 128 is mandatory for MiniMax MSA. KV cache is left at the default
-# dtype: amd/MiniMax-M3-MXFP4 ships no calibrated FP8 KV scales, so
-# --kv_cache_dtype fp8 trips an assertion (k_scale is None) in the MSA
-# fused_qknorm kernel during init.
+export AITER_QUICK_REDUCE_QUANTIZATION=INT4
+export AITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0
+export ATOM_M3_SPARSE_USE_ASM_PA=1
+export MAX_MODEL_LEN=32768
+export MAX_NUM_BATCHED_TOKENS=32768
+export MAX_NUM_SEQS=256
 python3 -m atom.entrypoints.openai_server \
     --model $MODEL \
     --server-port $PORT \
-    -tp $TP \
-    --max-model-len $MAX_MODEL_LEN $EP \
+    "${PARALLEL_ARGS[@]}" \
+    "${SPEC_ARGS[@]}" \
     --block-size 128 \
     --gpu-memory-utilization $MEM_FRAC_STATIC \
+    --max-model-len $MAX_MODEL_LEN \
+    --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
+    --max-num-seqs $MAX_NUM_SEQS \
+    --kv_cache_dtype fp8 \
     --trust-remote-code \
+    --no-enable_prefix_caching \
     > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
@@ -76,7 +75,7 @@ run_benchmark_serving \
     --max-concurrency "$CONC" \
     --result-filename "$RESULT_FILENAME" \
     --result-dir /workspace/ \
-    --trust-remote-code
+    --trust-remote-code $( [[ ${#SPEC_ARGS[@]} -gt 0 ]] && echo "--use-chat-template" )
 
 # After throughput, run evaluation only if RUN_EVAL is true
 if [ "${RUN_EVAL}" = "true" ]; then
@@ -86,4 +85,4 @@ fi
 
 # Stop GPU monitoring
 stop_gpu_monitor
-set +x
+set +x
\ No newline at end of file
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
new file mode 100644
index 000000000..fe71c47a0
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -0,0 +1,88 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    EP_SIZE \
+    DP_ATTENTION 
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
+
+SERVER_LOG=/workspace/server.log
+
+PARALLEL_ARGS=(-tp "$TP") #TP
+if [ "$DP_ATTENTION" = "true" ]; then
+    if [ "$EP_SIZE" -gt 1 ]; then #DP+EP
+        PARALLEL_ARGS=(-tp "$TP" --enable-expert-parallel --enable-dp-attention )
+    else #DP+TP
+        PARALLEL_ARGS=(-tp "$TP" --enable-dp-attention )
+    fi
+fi 
+
+SPEC_ARGS=(--method eagle3 --draft-model Inferact/MiniMax-M3-EAGLE3 --num-speculative-tokens 3 )
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+MEM_FRAC_STATIC=0.8
+
+set -x
+export AITER_QUICK_REDUCE_QUANTIZATION=INT4
+export AITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0
+export ATOM_M3_SPARSE_USE_ASM_PA=1
+export MAX_MODEL_LEN=32768
+export MAX_NUM_BATCHED_TOKENS=32768
+export MAX_NUM_SEQS=256
+python3 -m atom.entrypoints.openai_server \
+    --model $MODEL \
+    --server-port $PORT \
+    "${PARALLEL_ARGS[@]}" \
+    "${SPEC_ARGS[@]}" \
+    --block-size 128 \
+    --gpu-memory-utilization $MEM_FRAC_STATIC \
+    --max-model-len $MAX_MODEL_LEN \
+    --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
+    --max-num-seqs $MAX_NUM_SEQS \
+    --kv_cache_dtype fp8 \
+    --trust-remote-code \
+    --no-enable_prefix_caching \
+    > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+export PYTHONDONTWRITEBYTECODE=1
+run_benchmark_serving \
+    --model "$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 $( [[ ${#SPEC_ARGS[@]} -gt 0 ]] && echo "--use-chat-template" )
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
\ No newline at end of file
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index bb20c93a3..1651f1d10 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3984,6 +3984,15 @@
     - "Expand search space for minimaxm3-fp4-mi355x-atom: add TP2 and TP8 configurations, extend concurrency range to 256 for ISL1024 and ISL8192, and add TP8 conc=1-2 for ISL8192."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1825
 
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom
+    - minimaxm3-fp4-mi355x-atom-mtp
+  description:
+    - "Add minimaxm3-fp4-mi355x-atom-mtp: MiniMax-M3 MXFP4 on MI355X with EAGLE3 speculative decoding (3 draft tokens)"
+    - "Bump image to rocm/atom-dev:MiniMax-M3-20260623 for both fp4 atom entries"
+    - "Search space: TP2/TP4, ISL=1024,8192, OSL=1024, conc 1–256"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1917
+
 - config-keys:
     - minimaxm3-fp8-b300-vllm-mtp
   description: