From dddb199d2a03ca48adfd5aa1d0fc37ec555d8eb4 Mon Sep 17 00:00:00 2001
From: Patrick Kabwe <patrickckabwe@gmail.com>
Date: Thu, 10 Sep 2026 20:48:57 +0200
Subject: [PATCH] docs: document android xnnpack delegate

---
 README.md                  | 14 ++++++++++++++
 src/loadTensorflowModel.ts |  2 +-
 src/useTensorflowModel.ts  |  2 +-
 3 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index bece50df..02b6aad7 100644
--- a/README.md
+++ b/README.md
@@ -151,6 +151,20 @@ return <Camera device="back" isActive={true} outputs={[frameOutput]} {...otherPr
 > [!NOTE]
 > Unlike v4, VisionCamera v5 no longer requires boxing the model with `NitroModules.box()`. Since v5 is built on Nitro Modules and uses [react-native-worklets](https://docs.swmansion.com/react-native-worklets/), worklets can access HybridObjects like the TFLite model directly.
 
+### Using Android XNNPACK CPU Delegate
+
+Pass `['xnnpack']` to opt into the Android XNNPACK CPU delegate:
+
+```ts
+const model = await loadTensorflowModel(require('assets/my-model.tflite'), [
+  'xnnpack',
+])
+```
+
+Passing `[]` keeps the standard TFLite CPU path. XNNPACK is Android-only, uses an internally selected thread count capped for mobile devices, and cannot currently be combined with `android-gpu` or `nnapi`.
+
+Use XNNPACK when CPU inference is the right fit for your model or when GPU/NNAPI is unavailable or slower on your target devices. Benchmark your model on real devices, especially if the app also runs camera, image preprocessing, or other CPU-heavy work.
+
 ### Using GPU Delegates
 
 GPU Delegates offer faster, GPU-accelerated computation. There are multiple delegates available:
diff --git a/src/loadTensorflowModel.ts b/src/loadTensorflowModel.ts
index 04cf0b3e..f67593c2 100644
--- a/src/loadTensorflowModel.ts
+++ b/src/loadTensorflowModel.ts
@@ -22,7 +22,7 @@ const tfliteModule =
  * * If you are passing in a `{ url: ... }`, make sure the URL points directly to a `.tflite` model. This can either be a web URL (`http://..`/`https://..`), or a local file (`file://..`).
  *
  * @param source The `.tflite` model in form of either a `require(..)` statement or a `{ url: string }`.
- * @param delegates The delegates to use for computations. Uses the standard CPU delegate per default. The `core-ml` or `metal` delegates are GPU-accelerated, but don't work on every model.
+ * @param delegates The delegates to use for computations. Uses the standard CPU delegate per default. On Android, use `xnnpack` to opt into XNNPACK CPU acceleration. The `core-ml`, `metal`, `android-gpu`, and `nnapi` delegates are hardware-accelerated, but don't work on every model.
  * @returns The loaded Model.
  */
 export async function loadTensorflowModel(
diff --git a/src/useTensorflowModel.ts b/src/useTensorflowModel.ts
index d833142c..c97f8f15 100644
--- a/src/useTensorflowModel.ts
+++ b/src/useTensorflowModel.ts
@@ -24,7 +24,7 @@ export type TensorflowPlugin =
  * * If you are passing in a `{ url: ... }`, make sure the URL points directly to a `.tflite` model. This can either be a web URL (`http://..`/`https://..`), or a local file (`file://..`).
  *
  * @param source The `.tflite` model in form of either a `require(..)` statement or a `{ url: string }`.
- * @param delegates The delegates to use for computations. Uses the standard CPU delegate per default. The `core-ml` or `metal` delegates are GPU-accelerated, but don't work on every model.
+ * @param delegates The delegates to use for computations. Uses the standard CPU delegate per default. On Android, use `xnnpack` to opt into XNNPACK CPU acceleration. The `core-ml`, `metal`, `android-gpu`, and `nnapi` delegates are hardware-accelerated, but don't work on every model.
  * @returns The state of the Model.
  */
 export function useTensorflowModel(