diff --git a/docker/alphafold2.dockerfile b/docker/alphafold2.dockerfile
index 6d484dc5..a8ccc85c 100644
--- a/docker/alphafold2.dockerfile
+++ b/docker/alphafold2.dockerfile
@@ -64,6 +64,14 @@ RUN pip3 install --upgrade pip --no-cache-dir \
     && pip3 install --upgrade --no-cache-dir \
       "jax[cuda12]"==0.5.3
 
+# `pip install --upgrade "jax[cuda12]"==0.5.3` above drags in numpy 2.x, which makes
+# AlphaFold multimer fail at runtime: alphafold/data/msa_pairing.py does
+# `np.sum(x for x in feats)` and numpy>=2 raises
+# `TypeError: Calling np.sum(generator) is deprecated`. The numpy<2 pins earlier in
+# this file (conda + pyproject) run BEFORE the jax upgrade so they do not stick; re-pin
+# numpy<2 as the LAST dependency step. jax 0.5.3 runs fine with numpy 1.26.x.
+RUN pip install --no-cache-dir "numpy<2"
+
 # Strip Python caches to reduce layer size
 RUN find /opt/conda -type d -name "__pycache__" -prune -exec rm -rf {} + \
  && find /opt/conda -type f -name "*.pyc" -delete \