diff --git a/Cargo.lock b/Cargo.lock
index 3707c1c..13673bf 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1355,6 +1355,7 @@ dependencies = [
  "half",
  "indicatif",
  "native-tls",
+ "objc2",
  "ort",
  "serde",
  "serde_json",
diff --git a/Cargo.toml b/Cargo.toml
index 11b61fb..6fc784e 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -43,3 +43,5 @@ walkdir = "2"
 
 [target.'cfg(all(target_os = "macos", target_arch = "aarch64"))'.dependencies]
 coreml-native = "=0.2.0"
+# Remove once coreml-native includes an autorelease pool in predict_batch().
+objc2 = "0.6"
diff --git a/src/embed.rs b/src/embed.rs
index 13ef0ec..7b7a93f 100644
--- a/src/embed.rs
+++ b/src/embed.rs
@@ -513,7 +513,9 @@ impl EmbeddingEngine {
         let input_refs: Vec<&[(&str, &dyn AsMultiArray)]> =
             inputs.iter().map(Vec::as_slice).collect();
         let batch = BatchProvider::new(&input_refs)?;
-        let predictions = self.model.predict_batch(&batch)?;
+        // coreml-native 0.2.0 does not drain autoreleased CoreML batch
+        // outputs, eventually exhausting macOS's 16,384-IOSurface limit.
+        let predictions = objc2::rc::autoreleasepool(|_pool| self.model.predict_batch(&batch))?;
         (0..texts.len())
             .map(|i| {
                 let (embedding, shape) = predictions.get_f32(i, "div_1")?;