diff --git a/src/lerobot/configs/default.py b/src/lerobot/configs/default.py
index a03c436cf..f741fc6ae 100644
--- a/src/lerobot/configs/default.py
+++ b/src/lerobot/configs/default.py
@@ -98,12 +98,14 @@ class EMAConfig:
     Cost: 1× model params in fp32 shadow (~13 GB for pi052's 3.3B
     params) + one elementwise update per training step (~1% step time).
 
-    Off by default (back-compat). Recommended for long pi052 training
-    runs — typically ~1–3% absolute success-rate improvement on
-    closed-loop tasks per the diffusion-policy literature.
+    On by default — matches openpi (JAX) which ships EMA on for every
+    config, and closes the gap with the openpi PyTorch port which
+    explicitly lists EMA as unsupported. Set ``--ema.enable=false`` to
+    disable for short runs / memory-constrained training where the
+    extra fp32 shadow copy is the bottleneck.
     """
 
-    enable: bool = False
+    enable: bool = True
     # Target EMA decay β in θ_ema ← β·θ_ema + (1-β)·θ_live (passed to
     # ema-pytorch as ``beta``).
     #   0.999  — last ~1000 steps; pi05_libero default in openpi