diff --git a/tests/v1/engine/test_engine_args.py b/tests/v1/engine/test_engine_args.py
index c03c244bb1..5033f4768b 100644
--- a/tests/v1/engine/test_engine_args.py
+++ b/tests/v1/engine/test_engine_args.py
@@ -99,3 +99,20 @@ def test_mm_prefix_lm_raises_batched_tokens_floor():
         vllm_config = engine_args.create_engine_config(UsageContext.OPENAI_API_SERVER)
 
     assert vllm_config.scheduler_config.max_num_batched_tokens >= 2496
+
+
+def test_data_parallel_start_rank_zero_infers_hybrid_lb():
+    """An explicit --data-parallel-start-rank 0 must be treated the same as
+    any other explicit start rank when inferring hybrid LB mode, not as
+    "unset" (regression test for a truthiness-vs-`is not None` bug).
+    """
+    engine_args = EngineArgs(
+        model="facebook/opt-125m",
+        data_parallel_size=4,
+        data_parallel_size_local=2,
+        data_parallel_start_rank=0,
+    )
+    vllm_config = engine_args.create_engine_config(UsageContext.OPENAI_API_SERVER)
+
+    assert vllm_config.parallel_config.data_parallel_hybrid_lb is True
+    assert vllm_config.parallel_config.data_parallel_rank == 0
diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 4e4448a09c..528c884d73 100644
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -2158,7 +2158,7 @@ class EngineArgs:
         elif self.data_parallel_size_local is not None:
             data_parallel_size_local = self.data_parallel_size_local
 
-            if self.data_parallel_start_rank and not headless:
+            if self.data_parallel_start_rank is not None and not headless:
                 # Infer hybrid LB mode.
                 self.data_parallel_hybrid_lb = True
 
@@ -2177,7 +2177,9 @@ class EngineArgs:
                 self.data_parallel_hybrid_lb = False
 
             self.data_parallel_rank = (
-                self.data_parallel_start_rank or inferred_data_parallel_rank
+                self.data_parallel_start_rank
+                if self.data_parallel_start_rank is not None
+                else inferred_data_parallel_rank
             )
             if self.nnodes > 1:
                 logger.info(