diff --git a/langchain/src/chat_models/llama_cpp.ts b/langchain/src/chat_models/llama_cpp.ts
index 76ad790708c8..0a05745a0522 100644
--- a/langchain/src/chat_models/llama_cpp.ts
+++ b/langchain/src/chat_models/llama_cpp.ts
@@ -115,7 +115,7 @@ export class ChatLlamaCpp extends SimpleChatModel<LlamaCppCallOptions> {
   /** @ignore */
   async _call(
     messages: BaseMessage[],
-    _options: this["ParsedCallOptions"]
+    options: this["ParsedCallOptions"]
   ): Promise<string> {
     let prompt = "";
 
@@ -136,6 +136,7 @@ export class ChatLlamaCpp extends SimpleChatModel<LlamaCppCallOptions> {
 
     try {
       const promptOptions = {
+        onToken: options.onToken,
         maxTokens: this?.maxTokens,
         temperature: this?.temperature,
         topK: this?.topK,
diff --git a/langchain/src/llms/llama_cpp.ts b/langchain/src/llms/llama_cpp.ts
index 5081bde04008..f2d6518ffd2b 100644
--- a/langchain/src/llms/llama_cpp.ts
+++ b/langchain/src/llms/llama_cpp.ts
@@ -72,10 +72,11 @@ export class LlamaCpp extends LLM<LlamaCppCallOptions> {
   /** @ignore */
   async _call(
     prompt: string,
-    _options?: this["ParsedCallOptions"]
+    options?: this["ParsedCallOptions"]
   ): Promise<string> {
     try {
       const promptOptions = {
+        onToken: options?.onToken,
         maxTokens: this?.maxTokens,
         temperature: this?.temperature,
         topK: this?.topK,