diff --git a/searches/_runner.py b/searches/_runner.py
index 8eee5ad..2981a57 100644
--- a/searches/_runner.py
+++ b/searches/_runner.py
@@ -201,7 +201,8 @@ def _sampler_config_dict(
         return {
             "n_live": n_live,
             "num_mcmc_steps": int(_NSS_DEFAULTS["num_mcmc_steps"]),
-            "num_delete": min(int(_NSS_DEFAULTS["num_delete"]), batch),
+            "num_delete": int(_NSS_DEFAULTS["num_delete"]),
+            "chunk_size": batch,
             "termination": float(_NSS_DEFAULTS["termination"]),
             "seed": int(_NSS_DEFAULTS["seed"]),
             "jax_native": True,
diff --git a/searches/_samplers.py b/searches/_samplers.py
index 82a1e40..09ac2e0 100644
--- a/searches/_samplers.py
+++ b/searches/_samplers.py
@@ -171,21 +171,25 @@ def build_nss(
             "for the NumPy-front profile."
         )
     n_live = n_live_for(dataset_class, model_type)
-    # NSS's ``num_delete`` plays the same role as Nautilus ``n_batch``: it
-    # controls how many likelihoods fire in parallel per outer iteration.
-    # Cap it at the per-cell vmap budget so heavy cells (delaunay, inversion-
-    # based) don't OOM the A100. Floor at the default so small cells still
-    # benefit from sane batching.
-    num_delete = min(
-        int(_NSS_DEFAULTS["num_delete"]),
-        vmap_batch_for_cell(dataset_class, model_type, instrument),
-    )
+    # Memory-budget plumbing:
+    # - ``num_delete`` stays at the sampler's preferred default (50 particles
+    #   per outer iteration) so convergence isn't compromised.
+    # - ``chunk_size = vmap_batch_for_cell(...)`` caps the inner-vmap fan-out
+    #   per the A100-probed budget. PyAutoFit#1303 swaps blackjax's internal
+    #   ``jax.vmap(num_delete)`` for ``jax.lax.map(batch_size=chunk_size)``
+    #   when ``chunk_size < num_delete``, so peak GPU memory becomes
+    #   ``chunk_size × per_particle_state`` instead of ``num_delete × ...``.
+    # - ``chunk_size = None`` would also work on cells where the probe value
+    #   is >= num_delete (e.g. point_source fallback), but passing the probe
+    #   value explicitly keeps the JSON record honest about what was capped.
+    chunk_size = vmap_batch_for_cell(dataset_class, model_type, instrument)
     return af.NSS(
         name=config_name,
         path_prefix=f"searches/{sampler}/{dataset_class}/{model_type}/{instrument}",
         n_live=n_live,
         num_mcmc_steps=int(_NSS_DEFAULTS["num_mcmc_steps"]),
-        num_delete=num_delete,
+        num_delete=int(_NSS_DEFAULTS["num_delete"]),
+        chunk_size=chunk_size,
         termination=float(_NSS_DEFAULTS["termination"]),
         seed=int(_NSS_DEFAULTS["seed"]),
     )