diff --git a/crates/nexus_python3d/src/nexus.rs b/crates/nexus_python3d/src/nexus.rs
index 0cee52f..6a56091 100644
--- a/crates/nexus_python3d/src/nexus.rs
+++ b/crates/nexus_python3d/src/nexus.rs
@@ -282,6 +282,13 @@ impl NexusState {
         self.0.set_rbd_steps_per_frame(steps);
     }
 
+    /// Overrides the per-environment collision-pair capacity (default 4096)
+    /// used when the GPU state is allocated at `finalize`. Lower it for many
+    /// small batched envs; pair-keyed buffers scale as capacity × envs.
+    fn set_rbd_collisions_capacity(&mut self, capacity: u32) {
+        self.0.set_rbd_collisions_capacity(capacity);
+    }
+
     fn set_rbd_gravity(&mut self, viewer: PyRef<NexusViewer>, gravity: Vec3) {
         self.0
             .set_rbd_gravity(viewer.backend(), [gravity.0.x, gravity.0.y, gravity.0.z]);
diff --git a/src/state.rs b/src/state.rs
index e4dd821..625141e 100644
--- a/src/state.rs
+++ b/src/state.rs
@@ -160,6 +160,15 @@ impl NexusState {
         self.rbd_steps_per_frame = steps.max(1);
     }
 
+    /// Overrides the per-environment collision-pair capacity used when the
+    /// GPU rigid-body state is (re)allocated at `finalize`. The default (4096)
+    /// is sized for one busy scene, not thousands of small batched envs —
+    /// pair-keyed workspaces scale as `capacity x num_envs x sizeof(manifold)`,
+    /// which at 2048 envs binds ~9 GiB unless this is lowered.
+    pub fn set_rbd_collisions_capacity(&mut self, capacity: u32) {
+        self.capacities.rbd.collisions_capacity = capacity.max(1);
+    }
+
     /// Number of rigid-body solver steps per [`NexusPipeline::simulate`](crate::pipeline::NexusPipeline::simulate) call.
     pub fn rbd_steps_per_frame(&self) -> u32 {
         self.rbd_steps_per_frame