diff --git a/.runpod/hub.json b/.runpod/hub.json
index 671e489..279140b 100644
--- a/.runpod/hub.json
+++ b/.runpod/hub.json
@@ -23,7 +23,8 @@
           "TENSOR_PARALLEL_SIZE": "1",
           "GPU_MEMORY_UTILIZATION": "0.90",
           "MAX_CONCURRENCY": "30",
-          "QUANTIZATION": "fp8"
+          "QUANTIZATION": "fp8",
+          "MAX_NUM_SEQS": "32"
         }
       }
     ],
@@ -61,6 +62,17 @@
           "advanced": false
         }
       },
+      {
+        "key": "MAX_NUM_SEQS",
+        "input": {
+          "name": "Max Num Seqs",
+          "type": "string",
+          "description": "Most sequences vLLM schedules at once; Holo4's Mamba cache holds 125 on a 48 GB GPU",
+          "default": "32",
+          "required": false,
+          "advanced": true
+        }
+      },
       {
         "key": "MODEL_NAME",
         "input": {
diff --git a/.runpod/tests.json b/.runpod/tests.json
index 3ef0e83..9faeb72 100644
--- a/.runpod/tests.json
+++ b/.runpod/tests.json
@@ -33,6 +33,10 @@
         "key": "MAX_MODEL_LEN",
         "value": "8192"
       },
+      {
+        "key": "MAX_NUM_SEQS",
+        "value": "32"
+      },
       {
         "key": "MODEL_NAME",
         "value": "Hcompany/Holo4-35B-A3B"