diff --git a/src/lean_spec/subspecs/xmss/poseidon.py b/src/lean_spec/subspecs/xmss/poseidon.py
index 790314b0..3f1ef7a5 100644
--- a/src/lean_spec/subspecs/xmss/poseidon.py
+++ b/src/lean_spec/subspecs/xmss/poseidon.py
@@ -22,7 +22,7 @@
 
 from __future__ import annotations
 
-from pydantic import model_validator
+from pydantic import PrivateAttr, model_validator
 
 from lean_spec.types import StrictBaseModel
 
@@ -46,12 +46,25 @@ class PoseidonXmss(StrictBaseModel):
     params24: Poseidon1Params
     """Poseidon1 parameters for 24-width permutation."""
 
+    _engine16: Poseidon1 | None = PrivateAttr(default=None)
+    _engine24: Poseidon1 | None = PrivateAttr(default=None)
+
     @model_validator(mode="after")
     def _validate_strict_types(self) -> PoseidonXmss:
         """Reject subclasses to prevent type confusion attacks."""
         enforce_strict_types(self, params16=Poseidon1Params, params24=Poseidon1Params)
         return self
 
+    def _get_engine(self, width: int) -> Poseidon1:
+        """Return a cached Poseidon1 engine for the given width."""
+        if width == 16:
+            if self._engine16 is None:
+                self._engine16 = Poseidon1(self.params16)
+            return self._engine16
+        if self._engine24 is None:
+            self._engine24 = Poseidon1(self.params24)
+        return self._engine24
+
     def compress(self, input_vec: list[Fp], width: int, output_len: int) -> list[Fp]:
         """
         Implements the Poseidon1 hash in **compression mode**.
@@ -85,8 +98,7 @@ def compress(self, input_vec: list[Fp], width: int, output_len: int) -> list[Fp]
         # Select the correct permutation parameters based on the state width.
         if width not in (16, 24):
             raise ValueError(f"Width must be 16 or 24, got {width}")
-        params = self.params16 if width == 16 else self.params24
-        engine = Poseidon1(params)
+        engine = self._get_engine(width)
 
         # Create a padded input by extending with zeros to match the state width.
         padded_input = list(input_vec) + [Fp(value=0)] * (width - len(input_vec))
@@ -175,7 +187,7 @@ def sponge(
         # Determine the permutation parameters and the size of the rate.
         if width not in (16, 24):
             raise ValueError(f"Width must be 16 or 24, got {width}")
-        params = self.params16 if width == 16 else self.params24
+        engine = self._get_engine(width)
         rate = width - len(capacity_value)
 
         # Pad the input vector with zeros to be an exact multiple of the rate size.
@@ -189,9 +201,6 @@ def sponge(
         state = [Fp(value=0)] * width
         state[:cap_len] = capacity_value
 
-        # Create the engine once for efficiency.
-        engine = Poseidon1(params)
-
         # Absorb the input in rate-sized chunks via replacement.
         for i in range(0, len(padded_input), rate):
             chunk = padded_input[i : i + rate]