diff --git a/src/torchada/triton/autotune/fused_moe/configs/triton_3_2_0/E=768,N=1280,device_name=MTT_S5000.json b/src/torchada/triton/autotune/fused_moe/configs/triton_3_2_0/E=768,N=1280,device_name=MTT_S5000.json
new file mode 100644
index 0000000..32c7856
--- /dev/null
+++ b/src/torchada/triton/autotune/fused_moe/configs/triton_3_2_0/E=768,N=1280,device_name=MTT_S5000.json
@@ -0,0 +1,26 @@
+{
+  "21996": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 32,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 16,
+    "num_stages": 1
+  },
+  "24468": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 32,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 16,
+    "num_stages": 1
+  },
+  "45012": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 32,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 16,
+    "num_stages": 1
+  }
+}
diff --git a/src/torchada/triton/autotune/fused_moe/configs/triton_3_2_0/E=768,N=1280,device_name=MTT_S5000_down.json b/src/torchada/triton/autotune/fused_moe/configs/triton_3_2_0/E=768,N=1280,device_name=MTT_S5000_down.json
new file mode 100644
index 0000000..99f9bbc
--- /dev/null
+++ b/src/torchada/triton/autotune/fused_moe/configs/triton_3_2_0/E=768,N=1280,device_name=MTT_S5000_down.json
@@ -0,0 +1,26 @@
+{
+  "21996": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 32,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 16,
+    "num_stages": 1
+  },
+  "24468": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 32,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 16,
+    "num_stages": 1
+  },
+  "45012": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 16,
+    "num_stages": 1
+  }
+}
diff --git a/src/torchada/triton/autotune/fused_moe/configs/triton_3_6_0/E=768,N=1280,device_name=MTT_S5000,dtype=bf16.json b/src/torchada/triton/autotune/fused_moe/configs/triton_3_6_0/E=768,N=1280,device_name=MTT_S5000,dtype=bf16.json
new file mode 100644
index 0000000..ba6cb1c
--- /dev/null
+++ b/src/torchada/triton/autotune/fused_moe/configs/triton_3_6_0/E=768,N=1280,device_name=MTT_S5000,dtype=bf16.json
@@ -0,0 +1,10 @@
+{
+  "45012": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 32,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 16,
+    "num_stages": 1
+  }
+}
diff --git a/src/torchada/triton/runtime/fused_moe/config.py b/src/torchada/triton/runtime/fused_moe/config.py
index c34c3d6..19da803 100644
--- a/src/torchada/triton/runtime/fused_moe/config.py
+++ b/src/torchada/triton/runtime/fused_moe/config.py
@@ -71,7 +71,7 @@ def get_moe_configs(
     be picked and the associated configuration chosen to invoke the kernel.
     """
     # Supported Triton versions, should be sorted from the newest to the oldest
-    supported_triton_versions = ["3.4.0", "3.3.1", "3.2.0", "3.1.0"]
+    supported_triton_versions = ["3.6.0", "3.4.0", "3.3.1", "3.2.0", "3.1.0"]
 
     # First look up if an optimized configuration is available in the configs
     # directory
diff --git a/tests/test_magi2_s5000_moe_configs.py b/tests/test_magi2_s5000_moe_configs.py
new file mode 100644
index 0000000..45a5d9e
--- /dev/null
+++ b/tests/test_magi2_s5000_moe_configs.py
@@ -0,0 +1,32 @@
+import json
+from pathlib import Path
+
+CONFIG_DIR = (
+    Path(__file__).parents[1] / "src/torchada/triton/autotune/fused_moe/configs/triton_3_2_0"
+)
+NAMES = (
+    "E=768,N=1280,device_name=MTT_S5000.json",
+    "E=768,N=1280,device_name=MTT_S5000_down.json",
+)
+
+
+def test_magi2_s5000_configs_have_tuned_runtime_shapes():
+    for name in NAMES:
+        data = json.loads((CONFIG_DIR / name).read_text())
+        assert set(data) == {"21996", "24468", "45012"}
+        for shape in ("21996", "24468"):
+            assert data[shape] == _expected_config(block_size_k=32)
+
+        expected_k = 64 if name.endswith("_down.json") else 32
+        assert data["45012"] == _expected_config(block_size_k=expected_k)
+
+
+def _expected_config(*, block_size_k: int):
+    return {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": block_size_k,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 16,
+        "num_stages": 1,
+    }