diff --git a/saxml/server/pax/lm/params/template.py b/saxml/server/pax/lm/params/template.py
index a0ae3e9..4d31f90 100644
--- a/saxml/server/pax/lm/params/template.py
+++ b/saxml/server/pax/lm/params/template.py
@@ -56,6 +56,7 @@ class CommonServingTemplate:
   MAX_SEQ_LEN = None
   NUM_SAMPLES = 2
   TOP_K = 40
+  TEMPERATURE = 0.0
   TOP_K_RECALL_TARGET = 1.0  # When < 1.0, use tpu optimized approx_max_k
   USE_TOP_K_FOR_LOGPROBS = False
   BEAM_SIZE = 4
@@ -263,7 +264,7 @@ def generate(self) -> Optional[servable_lm_model.DecodeHParams]:
           max_decode_steps=self.MAX_DECODE_STEPS,
           seqlen=seqlen,
           num_samples=self.NUM_SAMPLES,
-          temperature=0.0,
+          temperature=self.TEMPERATURE,
           eos_id=stop_token_ids,
           k=self.TOP_K,
           top_k_recall_target=self.TOP_K_RECALL_TARGET,
@@ -347,7 +348,7 @@ def generate_stream(self) -> Optional[servable_lm_model.DecodeHParams]:
           max_decode_steps=self.MAX_DECODE_STEPS,
           seqlen=seqlen,
           num_samples=self.NUM_SAMPLES,
-          temperature=0.0,
+          temperature=self.TEMPERATURE,
           eos_id=stop_token_ids,
           k=self.TOP_K,
           top_k_recall_target=self.TOP_K_RECALL_TARGET,