react-native-nobodywho 5.0.0 → 5.1.0

package/android/src/main/AndroidManifest.xml

@@ -2,4 +2,7 @@
 <!-- Generated by uniffi-bindgen-react-native -->
 <manifest xmlns:android="http://schemas.android.com/apk/res/android"
           package="ooo.nobodywho">
-</manifest>
+    <application>
+        <uses-native-library android:name="libOpenCL.so" android:required="false" />
+    </application>
+</manifest>

package/android/src/main/AndroidManifestNew.xml

@@ -1,3 +1,6 @@
 <!-- Generated by uniffi-bindgen-react-native -->
 <manifest xmlns:android="http://schemas.android.com/apk/res/android">
+    <application>
+        <uses-native-library android:name="libOpenCL.so" android:required="false" />
+    </application>
 </manifest>

package/generated/ts/nobodywho.ts

@@ -152,7 +152,9 @@ export function getCachedModels(): Array<CachedModel> /*throws*/ {
  *
  * Pass `draft_model_path` pointing to a compatible MTP heads gguf (e.g.
  * `mtp-gemma-4-E2B-it.gguf` for Gemma-4-E2B) to enable MTP
- * speculative decoding on chats built from this model. Whether MTP is
+ * speculative decoding on chats built from this model. For models that
+ * bundle MTP layers (e.g. Qwen3.5), pass the model file itself; it is only
+ * loaded once. Whether MTP is
  * actually used is a per-chat decision — pass it through
  * `Chat`-level config on the wrapping binding.
  *
@@ -5168,7 +5170,7 @@ function uniffiEnsureInitialized() {
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_get_cached_models() !== 12002) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_func_get_cached_models");
     }
-    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_load_model() !== 22964) {
+    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_load_model() !== 8315) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_func_load_model");
     }
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_load_speech_to_text() !== 3224) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "react-native-nobodywho",
-  "version": "5.0.0",
+  "version": "5.1.0",
   "description": "Run LLMs locally and offline with React Native & Expo - Tool calling, RAG, Speech-to-Text, Text-to-Speech, and VAD",
   "main": "src/wrapper.ts",
   "types": "src/wrapper.ts",

package/src/model.ts

@@ -41,7 +41,8 @@ export class Model {
    * @param opts.projectionModelPath - Path to a multimodal projector file for vision models
    * @param opts.draftModelPath - Optional MTP draft-heads gguf. Loading it lets `Chat`
    *   instances opt into MTP speculative decoding via `new Chat({ ..., mtp: true })`.
-   *   Adds around 5% to VRAM usage.
+   *   Adds around 5% to VRAM usage. For models that bundle MTP layers (e.g. Qwen3.5),
+   *   pass the model file itself; it is only loaded once.
    * @param opts.onDownloadProgress - Invoked with `(downloaded, total)` byte counts while a remote model is being downloaded. Throttled to ~10 Hz with a guaranteed final emit on completion. Not invoked for cached/local files.
    */
   static async load(opts: {