diff --git a/.github/benchmark/oot_benchmark_models.json b/.github/benchmark/oot_benchmark_models.json
index 44c933461..7ae7e8d3e 100644
--- a/.github/benchmark/oot_benchmark_models.json
+++ b/.github/benchmark/oot_benchmark_models.json
@@ -164,7 +164,7 @@
           "prefix": "kimi-k25-mxfp4-tp4-met",
           "bench_args": "",
           "extra_args": "--trust-remote-code --tensor-parallel-size 4 --max-num-batched-tokens 16384 --max-model-len 16384",
-          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4"
+          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512"
         },
         {
           "tp_size": 4,
@@ -173,7 +173,7 @@
           "prefix": "kimi-k2-5-mxfp4-aw-tp4",
           "bench_args": "",
           "extra_args": "--trust-remote-code --tensor-parallel-size 4 --max-num-batched-tokens 16384 --max-model-len 16384",
-          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4"
+          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512"
         },
         {
           "tp_size": 8,
@@ -182,7 +182,7 @@
           "prefix": "kimi-k2-5-mxfp4-aw-tp8",
           "bench_args": "",
           "extra_args": "--trust-remote-code --tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-model-len 16384",
-          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4"
+          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512"
         }
       ]
     },
@@ -433,7 +433,7 @@
           "prefix": "minimax-m2-5-aw-tp2",
           "bench_args": "",
           "extra_args": "--trust-remote-code --tensor-parallel-size 2 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
-          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1"
+          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048"
         },
         {
           "tp_size": 4,
@@ -442,7 +442,7 @@
           "prefix": "minimax-m2-5-aw-tp4",
           "bench_args": "",
           "extra_args": "--trust-remote-code --tensor-parallel-size 4 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
-          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1"
+          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048"
         },
         {
           "tp_size": 8,
@@ -451,7 +451,7 @@
           "prefix": "minimax-m2-5-aw-tp8",
           "bench_args": "",
           "extra_args": "--trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
-          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1"
+          "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048"
         }
       ]
     }
diff --git a/.github/benchmark/oot_models_accuracy.json b/.github/benchmark/oot_models_accuracy.json
index f31486da6..ba957b413 100644
--- a/.github/benchmark/oot_models_accuracy.json
+++ b/.github/benchmark/oot_models_accuracy.json
@@ -165,7 +165,7 @@
     "model_name": "Kimi-K2.5-MXFP4 TP4",
     "model_path": "amd/Kimi-K2.5-MXFP4-AttnFP8",
     "extra_args": "--trust-remote-code --tensor-parallel-size 4 --max-num-batched-tokens 16384 --max-model-len 16384",
-    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4",
+    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512",
     "runner": "atom-plugin-acc-validation-runner",
     "test_level": "nightly",
     "priority": "P0",
@@ -175,7 +175,7 @@
     "model_name": "Kimi-K2.5-MXFP4 TP8",
     "model_path": "amd/Kimi-K2.5-MXFP4-AttnFP8",
     "extra_args": "--trust-remote-code --tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-model-len 16384",
-    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4",
+    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nAITER_AR_1STAGE_MAX_KB=512",
     "runner": "atom-plugin-acc-validation-runner",
     "test_level": "nightly",
     "priority": "P1",
@@ -310,7 +310,7 @@
     "model_name": "MiniMax-M2.5 TP2",
     "model_path": "MiniMaxAI/MiniMax-M2.5",
     "extra_args": "--trust-remote-code --tensor-parallel-size 2 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
-    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1",
+    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048",
     "runner": "atom-plugin-acc-validation-runner",
     "test_level": "nightly",
     "priority": "P0",
@@ -320,7 +320,7 @@
     "model_name": "MiniMax-M2.5 TP4",
     "model_path": "MiniMaxAI/MiniMax-M2.5",
     "extra_args": "--trust-remote-code --tensor-parallel-size 4 --kv-cache-dtype fp8 --max-num-batched-tokens 16384 --max-model-len 16384",
-    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1",
+    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nATOM_USE_GLUON_PA_DECODE=1\nAITER_AR_1STAGE_MAX_KB=2048",
     "runner": "atom-plugin-acc-validation-runner",
     "test_level": "nightly",
     "priority": "P1",
diff --git a/recipes/atom_vllm/Kimi-K2.5.md b/recipes/atom_vllm/Kimi-K2.5.md
index 81b9567ba..9ad34fb57 100644
--- a/recipes/atom_vllm/Kimi-K2.5.md
+++ b/recipes/atom_vllm/Kimi-K2.5.md
@@ -19,6 +19,10 @@ We adopt [amd/Kimi-K2.5-MXFP4-AttnFP8](https://huggingface.co/amd/Kimi-K2.5-MXFP
 ```bash
 # use quick allreduce to reduce TTFT
 export AITER_QUICK_REDUCE_QUANTIZATION=INT4
+# use flydsl MoE sorting kernel
+export AITER_USE_FLYDSL_MOE_SORTING=1
+# cap the single-stage allreduce size (KB) to favor the faster path
+export AITER_AR_1STAGE_MAX_KB=512
 
 vllm serve amd/Kimi-K2.5-MXFP4-AttnFP8 \
     --host localhost \
diff --git a/recipes/atom_vllm/MiniMax-M2.5.md b/recipes/atom_vllm/MiniMax-M2.5.md
index e421fcf15..1629a91ee 100644
--- a/recipes/atom_vllm/MiniMax-M2.5.md
+++ b/recipes/atom_vllm/MiniMax-M2.5.md
@@ -20,6 +20,8 @@ The following matches the vLLM-ATOM benchmark entries in `.github/benchmark/oot_
 export AITER_QUICK_REDUCE_QUANTIZATION=INT4
 export ATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1
 export ATOM_USE_GLUON_PA_DECODE=1
+# cap the single-stage allreduce size (KB) to favor the faster path
+export AITER_AR_1STAGE_MAX_KB=2048
 
 vllm serve MiniMaxAI/MiniMax-M2.5 \
     --host localhost \