diff --git a/CHANGELOG.md b/CHANGELOG.md
index d388cbf8..cc0d5829 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,7 @@
 ## Unreleased
 
+* Preserve accented text, emoji, and other Unicode characters when detokenizing native GGUF tokens.
+
 * Update native LiteRT-LM to `v0.17.0-5` with corrected Linux library loading, and enable explicit GPU selection on Linux x64 and Windows x64 while retaining CPU defaults.
 
 * Make zero-temperature native LiteRT-LM generation greedy to avoid corrupted GPU output.
diff --git a/lib/src/backends/llama_cpp/llama_cpp_service.dart b/lib/src/backends/llama_cpp/llama_cpp_service.dart
index 6a1acd41..73b86f2b 100644
--- a/lib/src/backends/llama_cpp/llama_cpp_service.dart
+++ b/lib/src/backends/llama_cpp/llama_cpp_service.dart
@@ -6019,7 +6019,8 @@ class LlamaCppService {
     final model = _models[modelHandle];
     if (model == null) return "";
     final vocab = llama_model_get_vocab(model.pointer);
-    final buffer = malloc<Int8>(256);
+    // UTF-8 decoding requires unsigned bytes, including byte-fallback tokens.
+    final buffer = malloc<Uint8>(256);
     final bytes = <int>[];
     for (final t in tokens) {
       final n = llama_token_to_piece(vocab, t, buffer.cast(), 256, 0, special);
diff --git a/test/integration/backends/llama_cpp/tokenizer_integration_test.dart b/test/integration/backends/llama_cpp/tokenizer_integration_test.dart
new file mode 100644
index 00000000..b36eac9d
--- /dev/null
+++ b/test/integration/backends/llama_cpp/tokenizer_integration_test.dart
@@ -0,0 +1,70 @@
+@TestOn('vm')
+@Timeout(Duration(minutes: 5))
+library;
+
+import 'package:llamadart/llamadart.dart';
+import 'package:test/test.dart';
+
+import '../../../test_helper.dart';
+
+void main() {
+  group('Native GGUF tokenizer', () {
+    late LlamaEngine engine;
+
+    setUpAll(() async {
+      final model = await TestHelper.getTestModel();
+      engine = LlamaEngine(LlamaBackend());
+      addTearDown(engine.dispose);
+      await engine.loadModel(
+        model.path,
+        modelParams: const ModelParams(
+          contextSize: 128,
+          gpuLayers: 0,
+          preferredBackend: GpuBackend.cpu,
+          numberOfThreads: 1,
+          numberOfThreadsBatch: 1,
+        ),
+      );
+    });
+
+    // The tiny Llama SentencePiece vocabulary adds one leading space. Keep
+    // that normalization explicit so trimming cannot hide damaged output.
+    const samples = {
+      'ASCII and newline': 'Hello world\n123',
+      'accented Latin': 'Montréal café',
+      'emoji': '👋',
+      'Korean': '한글',
+      'mixed Unicode': 'Montréal 👋\n한글 café',
+    };
+    for (final sample in samples.entries) {
+      test('round-trips ${sample.key}', () async {
+        final tokens = await engine.tokenize(sample.value, addSpecial: false);
+        expect(tokens, isNotEmpty);
+        expect(await engine.detokenize(tokens), ' ${sample.value}');
+      });
+    }
+
+    test('joins UTF-8 bytes across token boundaries before decoding', () async {
+      final tokens = await engine.tokenize('👋', addSpecial: false);
+      final pieces = await Future.wait(
+        tokens.map((token) => engine.detokenize([token])),
+      );
+
+      // This vocabulary emits individual byte tokens for this emoji. A lone
+      // partial sequence is malformed, but the full token sequence is valid.
+      expect(pieces.join(), contains('\uFFFD'));
+      expect(await engine.detokenize(tokens), ' 👋');
+    });
+
+    test('preserves special-token visibility', () async {
+      final tokens = await engine.tokenize('<s>', addSpecial: false);
+      expect(tokens, hasLength(1));
+      expect(await engine.detokenize(tokens), isEmpty);
+      expect(await engine.detokenize(tokens, special: true), '<s>');
+    });
+
+    test('detokenizes an empty token list', () async {
+      expect(await engine.detokenize([]), isEmpty);
+    });
+  });
+}
diff --git a/website/docs/changelog/recent-releases.md b/website/docs/changelog/recent-releases.md
index e64f2e05..98dbe3a5 100644
--- a/website/docs/changelog/recent-releases.md
+++ b/website/docs/changelog/recent-releases.md
@@ -9,6 +9,8 @@ For canonical full release notes, use:
 
 ## Unreleased
 
+* Preserve accented text, emoji, and other Unicode characters when detokenizing native GGUF tokens.
+
 * Update native LiteRT-LM to `v0.17.0-5` with corrected Linux library loading, and enable explicit GPU selection on Linux x64 and Windows x64 while retaining CPU defaults.
 
 * Make zero-temperature native LiteRT-LM generation greedy to avoid corrupted GPU output.