diff --git a/.github/benchmark/models.json b/.github/benchmark/models.json
index 863702228..f0a98a9c4 100644
--- a/.github/benchmark/models.json
+++ b/.github/benchmark/models.json
@@ -138,22 +138,52 @@
       "variants": [{ "label": "", "suffix": "", "conc_max": 256 }]
     },
     {
-      "display": "MiniMax-M2.7",
-      "path": "MiniMaxAI/MiniMax-M2.7",
-      "prefix": "MiniMax-M2.7",
+      "display": "MiniMax-M3-MXFP8",
+      "path": "MiniMaxAI/MiniMax-M3-MXFP8",
+      "prefix": "m3-mxfp8",
       "runner": "atom-mi355-8gpu.predownload",
-      "env_vars": "",
-      "config": { "tp": 2, "kv_cache_dtype": "fp8", "trust_remote_code": true },
-      "variants": [{ "label": "", "suffix": "", "conc_max": 256 }]
+      "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_FORCE_ATTN_TRITON=1\nAITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0",
+      "config": {
+        "tp": 4,
+        "kv_cache_dtype": "fp8",
+        "trust_remote_code": true,
+        "extra_args": "--gpu-memory-utilization 0.8 --block-size 128 --max-model-len 32768 --max-num-batched-tokens 32768 --no-enable_prefix_caching --online_quant_config '{\"global_quant_config\": \"ptpc_fp8\", \"exclude_layer\": [\"lm_head\", \"model.embed_tokens\", \"vision_tower\", \"multi_modal_projector\", \"patch_merge_mlp\", \"*block_sparse_moe\"]}'"
+      },
+      "variants": [
+        { "label": "", "suffix": "", "extra_args": "--max-num-seqs 256" },
+        {
+          "label": "EAGLE3",
+          "suffix": "-eagle3",
+          "extra_args": "--max-num-seqs 256 --method eagle3 --draft-model Inferact/MiniMax-M3-EAGLE3 --num-speculative-tokens 3",
+          "bench_args": "--use-chat-template",
+          "conc_min": 4,
+          "conc_max": 256
+        }
+      ]
     },
     {
-      "display": "MiniMax-M2.7-MXFP4",
-      "path": "amd/MiniMax-M2.7-MXFP4",
-      "prefix": "MiniMax-M2.7-MXFP4",
+      "display": "MiniMax-M3-MXFP4",
+      "path": "amd/MiniMax-M3-MXFP4",
+      "prefix": "m3-mxfp4",
       "runner": "atom-mi355-8gpu.predownload",
-      "env_vars": "",
-      "config": { "tp": 1, "kv_cache_dtype": "fp8", "trust_remote_code": true },
-      "variants": [{ "label": "", "suffix": "", "conc_max": 256 }]
+      "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_FORCE_ATTN_TRITON=1\nAITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0",
+      "config": {
+        "tp": 4,
+        "kv_cache_dtype": "fp8",
+        "trust_remote_code": true,
+        "extra_args": "--gpu-memory-utilization 0.8 --block-size 128 --max-model-len 32768 --max-num-batched-tokens 32768 --no-enable_prefix_caching"
+      },
+      "variants": [
+        { "label": "", "suffix": "", "extra_args": "--max-num-seqs 256" },
+        {
+          "label": "EAGLE3",
+          "suffix": "-eagle3",
+          "extra_args": "--max-num-seqs 256 --method eagle3 --draft-model Inferact/MiniMax-M3-EAGLE3 --num-speculative-tokens 3",
+          "bench_args": "--use-chat-template",
+          "conc_min": 4,
+          "conc_max": 256
+        }
+      ]
     },
     {
       "display": "Qwen3.5-397B-A17B-FP8",
diff --git a/.github/benchmark/models_accuracy.json b/.github/benchmark/models_accuracy.json
index e6c400faa..0f69b04fe 100644
--- a/.github/benchmark/models_accuracy.json
+++ b/.github/benchmark/models_accuracy.json
@@ -337,6 +337,32 @@
     "accuracy_baseline_model": "MiniMaxAI/MiniMax-M2.7",
     "_baseline_note": "ATOM CI measured BF16=0.9022 (gsm8k 3-shot flexible-extract). HF amd/MiniMax-M2.7-MXFP4: MXFP4=91.89, baseline=91.81 (percentage)."
   },
+  {
+    "model_name": "MiniMax-M3-MXFP4",
+    "model_path": "amd/MiniMax-M3-MXFP4",
+    "extraArgs": "--kv_cache_dtype fp8 -tp 8 --trust-remote-code --gpu-memory-utilization 0.8 --block-size 128 --max-model-len 32768 --max-num-seqs 128 --max-num-batched-tokens 32768 --no-enable_prefix_caching",
+    "client_command": "lm_eval --model local-chat-completions --apply_chat_template --fewshot_as_multiturn --model_args model=${MODEL_PATH},base_url=http://localhost:8000/v1/chat/completions,num_concurrent=32,max_retries=3,max_gen_toks=16384,tokenized_requests=False,trust_remote_code=True --tasks gsm8k --num_fewshot 5 --output_path ${OUTPUT_PATH}",
+    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_FORCE_ATTN_TRITON=1\nAITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0",
+    "runner": "linux-atom-do-mi350x-8",
+    "test_level": "pr",
+    "accuracy_threshold": 0.93,
+    "accuracy_baseline": 0.9363,
+    "accuracy_baseline_model": "amd/MiniMax-M3-MXFP4",
+    "_baseline_note": "FP4 M3 tp8. GSM8K 5-shot chat (apply_chat_template + fewshot_as_multiturn, num_concurrent=32, max_gen_toks=16384) per /data/users/zejun/zejun_atom/m3.md: validated MXFP4 flexible-extract=0.9363 ± 0.0067 (measured at tp4). Threshold 0.93 leaves ~0.6pp headroom; refresh after first CI measurement at tp8."
+  },
+  {
+    "model_name": "MiniMax-M3-MXFP4 Eagle3",
+    "model_path": "amd/MiniMax-M3-MXFP4",
+    "extraArgs": "--kv_cache_dtype fp8 -tp 8 --trust-remote-code --gpu-memory-utilization 0.8 --block-size 128 --max-model-len 32768 --max-num-seqs 256 --max-num-batched-tokens 32768 --no-enable_prefix_caching --method eagle3 --draft-model Inferact/MiniMax-M3-EAGLE3 --num-speculative-tokens 3",
+    "client_command": "lm_eval --model local-chat-completions --apply_chat_template --fewshot_as_multiturn --model_args model=${MODEL_PATH},base_url=http://localhost:8000/v1/chat/completions,num_concurrent=32,max_retries=3,max_gen_toks=16384,tokenized_requests=False,trust_remote_code=True --tasks gsm8k --num_fewshot 5 --output_path ${OUTPUT_PATH}",
+    "env_vars": "AITER_QUICK_REDUCE_QUANTIZATION=INT4\nATOM_FORCE_ATTN_TRITON=1\nAITER_QUICK_REDUCE_CAST_BF16_TO_FP16=0",
+    "runner": "linux-atom-do-mi350x-8",
+    "test_level": "nightly",
+    "accuracy_threshold": 0.93,
+    "accuracy_baseline": 0.9469,
+    "accuracy_baseline_model": "amd/MiniMax-M3-MXFP4 + Inferact/MiniMax-M3-EAGLE3",
+    "_baseline_note": "FP4 M3 + EAGLE3 draft (tp8), lossless vs greedy target. GSM8K 5-shot chat per /data/users/zejun/zejun_atom/m3.md: validated fp4_eagle_tp4 flexible-extract=0.9469 ± 0.0062 (accept ratio 73.36%, avg 3.20 toks/fwd, commit 9fc48338); baseline carried over from the tp4 validation since EAGLE3 is lossless wrt the target. Threshold 0.93 leaves ~1.7pp headroom; refresh after first CI measurement at tp8."
+  },
   {
     "model_name": "MiMo-V2-Flash",
     "model_path": "XiaomiMiMo/MiMo-V2-Flash",
diff --git a/.github/workflows/atom-benchmark.yaml b/.github/workflows/atom-benchmark.yaml
index 5fec695b2..96a1b385b 100644
--- a/.github/workflows/atom-benchmark.yaml
+++ b/.github/workflows/atom-benchmark.yaml
@@ -38,12 +38,12 @@ on:
         description: "Benchmark Kimi-K2.5-MXFP4"
         type: boolean
         default: true
-      MiniMax-M2.7:
-        description: "Benchmark MiniMax-M2.7"
+      m3-mxfp8:
+        description: "Benchmark MiniMax-M3-MXFP8 (+ EAGLE3)"
         type: boolean
         default: true
-      MiniMax-M2.7-MXFP4:
-        description: "Benchmark MiniMax-M2.7-MXFP4"
+      m3-mxfp4:
+        description: "Benchmark MiniMax-M3-MXFP4 (+ EAGLE3)"
         type: boolean
         default: true
       qwen35-397b-fp8: