diff --git a/skeliner/io.py b/skeliner/io.py
index 0d42552..6a106c5 100644
--- a/skeliner/io.py
+++ b/skeliner/io.py
@@ -355,8 +355,6 @@ def to_npz(
     if not path.suffix:
         path = path.with_suffix(".npz")
 
-    c = {} if not compress else {"compress": True}
-
     if cache_kdtree:
         skeleton._ensure_nodes_kdtree()
         skeleton._ensure_node_neighbors()
@@ -367,11 +365,39 @@ def to_npz(
     # ragged node2verts  → index + offset
     if skeleton.node2verts is not None:
         n2v_idx = np.concatenate(skeleton.node2verts)
-        n2v_off = np.cumsum([0, *map(len, skeleton.node2verts)]).astype(np.int64)
+        n2v_off = np.cumsum([0, *map(len, skeleton.node2verts)])
+
+        # Determine smallest dtype for n2v_idx
+        if len(n2v_idx) > 0:
+            idx_min, idx_max = n2v_idx.min(), n2v_idx.max()
+            if idx_min >= 0 and idx_max <= np.iinfo(np.uint16).max:
+                n2v_idx = n2v_idx.astype(np.uint16)
+            elif idx_min < 0 and idx_min >= np.iinfo(np.int16).min and idx_max <= np.iinfo(np.int16).max:
+                n2v_idx = n2v_idx.astype(np.int16)
+            elif idx_min >= 0 and idx_max <= np.iinfo(np.uint32).max:
+                n2v_idx = n2v_idx.astype(np.uint32)
+            elif idx_min < 0 and idx_min >= np.iinfo(np.int32).min and idx_max <= np.iinfo(np.int32).max:
+                n2v_idx = n2v_idx.astype(np.int32)
+            else:
+                n2v_idx = n2v_idx.astype(np.int64)
+        else:
+            n2v_idx = n2v_idx.astype(np.uint16)
+
+        # Determine smallest dtype for n2v_off
+        off_min, off_max = n2v_off.min(), n2v_off.max()
+        if off_min >= 0 and off_max <= np.iinfo(np.uint16).max:
+            n2v_off = n2v_off.astype(np.uint16)
+        elif off_min < 0 and off_min >= np.iinfo(np.int16).min and off_max <= np.iinfo(np.int16).max:
+            n2v_off = n2v_off.astype(np.int16)
+        elif off_min >= 0 and off_max <= np.iinfo(np.uint32).max:
+            n2v_off = n2v_off.astype(np.uint32)
+        elif off_min < 0 and off_min >= np.iinfo(np.int32).min and off_max <= np.iinfo(np.int32).max:
+            n2v_off = n2v_off.astype(np.int32)
+        else:
+            n2v_off = n2v_off.astype(np.int64)
     else:
-        n2v_idx = np.array([], dtype=np.int64)
-        n2v_off = np.array([0], dtype=np.int64)
-
+        n2v_idx = np.array([], dtype=np.uint16)
+        n2v_off = np.array([0], dtype=np.uint16)
     # ----------- NEW: persist the metadata dict -------------------------
     # We wrap it in a 0-D object array because np.savez can only store
     # ndarrays — this keeps the archive a single *.npz* with no sidecars.
@@ -397,7 +423,9 @@ def to_npz(
         tree_payload["neighbors_idx"] = data.astype(np.int64, copy=False)
         tree_payload["neighbors_off"] = offsets.astype(np.int64, copy=False)
 
-    np.savez(
+    save_fun = np.savez_compressed if compress else np.savez
+
+    save_fun(
         path,
         nodes=skeleton.nodes,
         edges=skeleton.edges,
@@ -416,7 +444,6 @@ def to_npz(
         **extra,
         **meta,
         **tree_payload,
-        **c,
     )