diff --git a/android/app/src/main/java/com/pocketpalai/MainApplication.kt b/android/app/src/main/java/com/pocketpalai/MainApplication.kt
index 28a7d6429..2231987a1 100644
--- a/android/app/src/main/java/com/pocketpalai/MainApplication.kt
+++ b/android/app/src/main/java/com/pocketpalai/MainApplication.kt
@@ -1,6 +1,7 @@
 package com.pocketpal
 
 import android.app.Application
+import android.system.Os
 import com.facebook.react.PackageList
 import com.facebook.react.ReactApplication
 import com.facebook.react.ReactHost
@@ -43,6 +44,12 @@ class MainApplication : Application(), ReactApplication {
 
   override fun onCreate() {
     super.onCreate()
+    // Enable Adreno large buffer support on Qualcomm A7X/A8X GPUs.
+    // The OpenCL backend in llama.rn self-gates on GPU family and the
+    // cl_qcom_large_buffer extension — this is a no-op on non-Adreno devices.
+    // Must be set before SoLoader.init so the native library picks it up.
+    // See: https://github.com/ggml-org/llama.cpp/pull/20997
+    Os.setenv("LM_GGML_OPENCL_ADRENO_USE_LARGE_BUFFER", "1", true)
     SoLoader.init(this, OpenSourceMergedSoMapping)
     if (BuildConfig.IS_NEW_ARCHITECTURE_ENABLED) {
       // If you opted-in for the New Architecture, we load the native entry point for this app.