diff --git a/LLama/LLamaExecutorBase.cs b/LLama/LLamaExecutorBase.cs
index 578bd4d84..f047ab892 100644
--- a/LLama/LLamaExecutorBase.cs
+++ b/LLama/LLamaExecutorBase.cs
@@ -70,6 +70,8 @@ public abstract class StatefulExecutorBase : ILLamaExecutor
         /// </summary>
         protected float? MirostatMu { get; set; }
 
+        private StreamingTokenDecoder _decoder;
+
         /// <summary>
         /// 
         /// </summary>
@@ -83,6 +85,7 @@ protected StatefulExecutorBase(LLamaContext context, ILogger? logger = null)
             _consumedTokensCount = 0;
             _n_session_consumed = 0;
             _last_n_tokens = new FixedSizeQueue<llama_token>(Context.ContextSize).FillWith(0);
+            _decoder = new StreamingTokenDecoder(context);
         }
 
         /// <summary>
@@ -294,7 +297,10 @@ public virtual async IAsyncEnumerable<string> InferAsync(string text, IInference
                 await InferInternal(inferenceParams, args);
 
                 if (args.ReturnValue)
-                    yield return Context.DeTokenize(_embeds);
+                {
+                    _decoder.AddRange(_embeds);
+                    yield return _decoder.Read();
+                }
 
                 var (breakGeneration, extraOutputs) = await PostProcess(inferenceParams, args);
                 if (extraOutputs is { Count: > 0 })