diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index 26bde59696..e7bdf27542 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -2039,6 +2039,9 @@ build_replay_cmd() {
     # CPU on minimax-m2.5 at high concurrency. Lossless for vLLM (server
     # usage is authoritative).
     REPLAY_CMD+=" --use-server-token-count"
+    if [ -n "${AIPERF_EXTRA_INPUTS:-}" ]; then
+        REPLAY_CMD+=" --extra-inputs $AIPERF_EXTRA_INPUTS"
+    fi
     # Dynamo's KV router needs an explicit conversation session binding to
     # keep later turns on the prefill worker that owns their prefix blocks.
     # X-Correlation-ID is useful tracing metadata but does not establish that
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-agentic.yaml
new file mode 100644
index 0000000000..4686d003fc
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-agentic.yaml
@@ -0,0 +1,80 @@
+name: "minimax-m3-vllm-agg-gb200-dep4-agentic"
+
+model: { path: "minimax-m3-nvfp4", container: "vllm/vllm-openai:v0.27.1", precision: "fp4" }
+identity:
+  model: { repo: "nvidia/MiniMax-M3-NVFP4" }
+  container: { image: "vllm/vllm-openai:v0.27.1" }
+  frameworks: { dynamo: "1.3.1" }
+dynamo: { version: "1.3.1", install: true }
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "12:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+resources: { gpu_type: "gb200", gpus_per_node: 4, agg_nodes: 1, agg_workers: 1, gpus_per_agg: 4 }
+infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 }
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env: { DYN_TCP_REQUEST_TIMEOUT: "60" }
+  args: { router-mode: "kv", router-kv-events: true, router-reset-states: true, router-temperature: "0", router-session-affinity-ttl-secs: 14400, kv-cache-block-size: 128 }
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  kv_events_config: { aggregated: true }
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "cuda_copy,cuda_ipc,rc"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  vllm_config:
+    aggregated:
+      served-model-name: "nvidia/MiniMax-M3-NVFP4"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}'
+      block-size: 128
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
+      stream-interval: 20
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 16384
+      no-enable-flashinfer-autotune: true
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_EXTRA_INPUTS: "thinking:true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-vllm-simple-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-vllm-simple-agentic.yaml
new file mode 100644
index 0000000000..528e08b835
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-vllm-simple-agentic.yaml
@@ -0,0 +1,82 @@
+name: "minimax-m3-vllm-agg-gb200-dep4-vllm-simple-agentic"
+
+model: { path: "minimax-m3-nvfp4", container: "vllm/vllm-openai:v0.27.1", precision: "fp4" }
+identity:
+  model: { repo: "nvidia/MiniMax-M3-NVFP4" }
+  container: { image: "vllm/vllm-openai:v0.27.1" }
+  frameworks: { dynamo: "1.3.1" }
+dynamo: { version: "1.3.1", install: true }
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "12:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+resources: { gpu_type: "gb200", gpus_per_node: 4, agg_nodes: 1, agg_workers: 1, gpus_per_agg: 4 }
+infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 }
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env: { DYN_TCP_REQUEST_TIMEOUT: "60" }
+  args: { router-mode: "kv", router-kv-events: true, router-reset-states: true, router-temperature: "0", router-session-affinity-ttl-secs: 14400, kv-cache-block-size: 128 }
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  kv_events_config: { aggregated: true }
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_SIMPLE_KV_OFFLOAD: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "cuda_copy,cuda_ipc,rc"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  vllm_config:
+    aggregated:
+      served-model-name: "nvidia/MiniMax-M3-NVFP4"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}'
+      block-size: 128
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
+      kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":549755813888,"cpu_bytes_to_use_per_rank":137438953472,"lazy_offload":false}}'
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
+      stream-interval: 20
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 16384
+      no-enable-flashinfer-autotune: true
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_EXTRA_INPUTS: "thinking:true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep8-agentic.yaml
new file mode 100644
index 0000000000..b8b4c33411
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep8-agentic.yaml
@@ -0,0 +1,80 @@
+name: "minimax-m3-vllm-agg-gb200-dep8-agentic"
+
+model: { path: "minimax-m3-nvfp4", container: "vllm/vllm-openai:v0.27.1", precision: "fp4" }
+identity:
+  model: { repo: "nvidia/MiniMax-M3-NVFP4" }
+  container: { image: "vllm/vllm-openai:v0.27.1" }
+  frameworks: { dynamo: "1.3.1" }
+dynamo: { version: "1.3.1", install: true }
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "12:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+resources: { gpu_type: "gb200", gpus_per_node: 4, agg_nodes: 2, agg_workers: 1, gpus_per_agg: 8 }
+infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 }
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env: { DYN_TCP_REQUEST_TIMEOUT: "60" }
+  args: { router-mode: "kv", router-kv-events: true, router-reset-states: true, router-temperature: "0", router-session-affinity-ttl-secs: 14400, kv-cache-block-size: 128 }
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  kv_events_config: { aggregated: true }
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "cuda_copy,cuda_ipc,rc"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  vllm_config:
+    aggregated:
+      served-model-name: "nvidia/MiniMax-M3-NVFP4"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}'
+      block-size: 128
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
+      stream-interval: 20
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 16384
+      no-enable-flashinfer-autotune: true
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_EXTRA_INPUTS: "thinking:true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-agentic.yaml
new file mode 100644
index 0000000000..b74dd665b4
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-agentic.yaml
@@ -0,0 +1,99 @@
+name: "minimax-m3-vllm-agg-gb200-tp4-agentic"
+
+model:
+  path: "minimax-m3-nvfp4"
+  container: "vllm/vllm-openai:v0.27.1"
+  precision: "fp4"
+
+identity:
+  model: { repo: "nvidia/MiniMax-M3-NVFP4" }
+  container: { image: "vllm/vllm-openai:v0.27.1" }
+  frameworks: { dynamo: "1.3.1" }
+
+dynamo: { version: "1.3.1", install: true }
+environment: { ETCD_LEASE_TTL: "7200" }
+
+slurm: { time_limit: "12:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  agg_nodes: 1
+  agg_workers: 1
+  gpus_per_agg: 4
+
+infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 }
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env: { DYN_TCP_REQUEST_TIMEOUT: "60" }
+  args:
+    router-mode: "kv"
+    router-kv-events: true
+    router-reset-states: true
+    router-temperature: "0"
+    router-session-affinity-ttl-secs: 14400
+    kv-cache-block-size: 128
+
+backend:
+  type: vllm
+  connector: null
+  kv_events_config: { aggregated: true }
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "cuda_copy,cuda_ipc,rc"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  vllm_config:
+    aggregated:
+      served-model-name: "nvidia/MiniMax-M3-NVFP4"
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}'
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
+      block-size: 128
+      gpu-memory-utilization: 0.9
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
+      stream-interval: 20
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 16384
+      no-enable-flashinfer-autotune: true
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_EXTRA_INPUTS: "thinking:true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml
new file mode 100644
index 0000000000..1d62599edd
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml
@@ -0,0 +1,87 @@
+name: "minimax-m3-vllm-agg-gb200-tp4-vllm-simple-agentic"
+
+model:
+  path: "minimax-m3-nvfp4"
+  container: "vllm/vllm-openai:v0.27.1"
+  precision: "fp4"
+
+identity:
+  model: { repo: "nvidia/MiniMax-M3-NVFP4" }
+  container: { image: "vllm/vllm-openai:v0.27.1" }
+  frameworks: { dynamo: "1.3.1" }
+
+dynamo: { version: "1.3.1", install: true }
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "12:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+resources: { gpu_type: "gb200", gpus_per_node: 4, agg_nodes: 1, agg_workers: 1, gpus_per_agg: 4 }
+infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 }
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env: { DYN_TCP_REQUEST_TIMEOUT: "60" }
+  args: { router-mode: "kv", router-kv-events: true, router-reset-states: true, router-temperature: "0", router-session-affinity-ttl-secs: 14400, kv-cache-block-size: 128 }
+
+backend:
+  type: vllm
+  connector: null
+  kv_events_config: { aggregated: true }
+  aggregated_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_SIMPLE_KV_OFFLOAD: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "cuda_copy,cuda_ipc,rc"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  vllm_config:
+    aggregated:
+      served-model-name: "nvidia/MiniMax-M3-NVFP4"
+      tensor-parallel-size: 4
+      pipeline-parallel-size: 1
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}'
+      block-size: 128
+      gpu-memory-utilization: 0.9
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
+      kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":549755813888,"cpu_bytes_to_use_per_rank":137438953472,"lazy_offload":false}}'
+      stream-interval: 20
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 16384
+      no-enable-flashinfer-autotune: true
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_EXTRA_INPUTS: "thinking:true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml
new file mode 100644
index 0000000000..7c2343f404
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml
@@ -0,0 +1,117 @@
+name: "minimax-m3-vllm-disagg-gb200-1p1d-dep8-dep4-agentic"
+
+model: { path: "minimax-m3-nvfp4", container: "vllm/vllm-openai:v0.27.1", precision: "fp4" }
+identity:
+  model: { repo: "nvidia/MiniMax-M3-NVFP4" }
+  container: { image: "vllm/vllm-openai:v0.27.1" }
+  frameworks: { dynamo: "1.3.1" }
+dynamo: { version: "1.3.1", install: true }
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "12:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 4
+infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 }
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env: { DYN_TCP_REQUEST_TIMEOUT: "60" }
+  args: { router-mode: "kv", router-kv-events: true, router-reset-states: true, router-temperature: "0", router-session-affinity-ttl-secs: 14400, kv-cache-block-size: 128 }
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  kv_events_config: true
+  prefill_environment: &worker_env
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "cuda_copy,cuda_ipc,rc"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  decode_environment: *worker_env
+  vllm_config:
+    prefill:
+      served-model-name: "nvidia/MiniMax-M3-NVFP4"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      kv-transfer-config: '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+      attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}'
+      block-size: 128
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
+      compilation-config: '{"cudagraph_mode":"PIECEWISE"}'
+      stream-interval: 20
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 16384
+      no-enable-flashinfer-autotune: true
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+    decode:
+      served-model-name: "nvidia/MiniMax-M3-NVFP4"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      kv-transfer-config: '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+      attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}'
+      block-size: 128
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
+      stream-interval: 20
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 16384
+      no-enable-flashinfer-autotune: true
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_EXTRA_INPUTS: "thinking:true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index b469673b45..8fd0430491 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7246,6 +7246,119 @@ minimaxm3-fp4-b200-vllm-agentic-mtp:
       search-space:
       - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 5, 8, 10, 12, 15, 20] }
       - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [20, 30, 40] }
+# Preserve the B200 TP4 search space and add the GB200 Pareto candidates found
+# by direct DEP and P/D tuning.
+minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp:
+  image: vllm/vllm-openai:v0.27.1
+  model: nvidia/MiniMax-M3-NVFP4
+  model-prefix: minimaxm3
+  runner: cluster:gb200-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.1" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.61
+      search-space:
+      - spec-decoding: mtp
+        kv-offloading: none
+        conc-list: [1, 2, 5, 8, 10, 12, 15, 20]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-agentic.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78"
+        decode: { num-worker: 0, tp: 4, ep: 1, dp-attn: false }
+      - spec-decoding: mtp
+        kv-offloading: dram
+        kv-offload-backend: { name: vllm-simple }
+        conc-list: [20, 30, 40]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78"
+        decode: { num-worker: 0, tp: 4, ep: 1, dp-attn: false }
+      - spec-decoding: mtp
+        kv-offloading: none
+        conc-list: [4, 32]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-agentic.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78"
+        decode: { num-worker: 0, tp: 4, ep: 4, dp-attn: true }
+      - spec-decoding: mtp
+        kv-offloading: dram
+        kv-offload-backend: { name: vllm-simple }
+        conc-list: [32, 40]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-vllm-simple-agentic.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78"
+        decode: { num-worker: 0, tp: 4, ep: 4, dp-attn: true }
+      - spec-decoding: mtp
+        kv-offloading: none
+        conc-list: [48]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep8-agentic.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78"
+        decode: { num-worker: 0, tp: 8, ep: 8, dp-attn: true }
+
+minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp:
+  image: vllm/vllm-openai:v0.27.1
+  model: nvidia/MiniMax-M3-NVFP4
+  model-prefix: minimaxm3
+  runner: cluster:gb200-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.1" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.61
+      search-space:
+      - spec-decoding: mtp
+        kv-offloading: none
+        conc-list: [48]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78"
+        decode: { num-worker: 1, tp: 4, ep: 4, dp-attn: true }
+
 dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-agg:
   image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
   model: deepseek-ai/DeepSeek-V4-Pro
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 067b8a79ff..bc516a3357 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5981,3 +5981,13 @@
     - "Inject the committed synthetic MTP acceptance length only for GB300 AgentX throughput; keep eval-only jobs on real target verification."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2571
   
+
+- config-keys:
+    - minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp
+    - minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add the tuned GB200 MiniMax-M3 FP4 AgentX frontier with EAGLE3, the B200 TP4 baseline, DEP4/DEP8, SimpleCPU offload, and KV-routed P/D."
+    - "Use full-decode-only CUDA graphs for TP4 stability."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2609
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index ed4cbed78b..c4fb8899b3 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -251,8 +251,11 @@ elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
     elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/MiniMax-M3-MXFP8"
         export SRT_SLURM_MODEL_PREFIX="minimax-m3-mxfp8"
+    elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" ]]; then
+        export MODEL_PATH="/mnt/lustre01/models/MiniMax-M3-NVFP4"
+        export SRT_SLURM_MODEL_PREFIX="minimax-m3-nvfp4"
     else
-        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, kimik3/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8, minimaxm3/fp8"
+        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, kimik3/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8, minimaxm3/fp4, minimaxm3/fp8"
         exit 1
     fi
 else
@@ -391,9 +394,21 @@ if [ -d "$SRT_REPO_DIR" ]; then
     rm -rf "$SRT_REPO_DIR"
 fi
 
+# MiniMax-M3 FP4 AgentX uses v1.0.50 for complete logical-worker metrics
+# discovery across aggregate, DP-attention, and disaggregated topologies.
+if [[ "$IS_AGENTIC" == "1" && "$MODEL_PREFIX" == "minimaxm3" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-vllm" ]]; then
+    git clone --branch v1.0.50 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR"
+    test "$(git rev-parse HEAD)" = "e4019633c9e2bc25f38c44b81edf52bb0504d937" || {
+        echo "Error: NVIDIA/srt-slurm v1.0.50 resolved to an unexpected commit" >&2
+        exit 1
+    }
+    mkdir -p recipes/vllm/minimax-m3/gb200-fp4/agentic
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic" \
+        recipes/vllm/minimax-m3/gb200-fp4/agentic
 # These AgentX submissions use released srt-slurm custom-benchmark metrics
 # discovery so AIPerf receives every logical worker endpoint.
-if [[ "$IS_AGENTIC" == "1" && (( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-sglang" ) || ( "$MODEL_PREFIX" == "dsv4" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-vllm" )) ]]; then
+elif [[ "$IS_AGENTIC" == "1" && (( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-sglang" ) || ( "$MODEL_PREFIX" == "dsv4" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-vllm" )) ]]; then
     git clone --branch v1.0.45 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
     test "$(git rev-parse HEAD)" = "9d8d92b20c350a5d42f0709f5a0b64e30eb37d33" || {