diff --git a/Cargo.lock b/Cargo.lock
index 0f7ce122..c7c5f3b9 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -281,6 +281,15 @@ version = "0.3.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a357d28ed41a50f9c765dbfe56cbc04a64e53e5fc58ba79fbc34c10ef3df831f"
 
+[[package]]
+name = "encoding_rs"
+version = "0.8.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7268b386296a025e474d5140678f75d6de9493ae55a5d709eeb9dd08149945e1"
+dependencies = [
+ "cfg-if",
+]
+
 [[package]]
 name = "errno"
 version = "0.3.8"
@@ -907,6 +916,7 @@ version = "0.1.46"
 dependencies = [
  "anyhow",
  "clap",
+ "encoding_rs",
  "hf-hub",
  "llama-cpp-2",
 ]
diff --git a/Cargo.toml b/Cargo.toml
index 986f8ee3..a81be6cc 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -19,6 +19,7 @@ bindgen = "0.69.4"
 cc = "1.0.90"
 anyhow = "1.0.81"
 clap = "4.5.4"
+encoding_rs = "0.8.33"
 
 [workspace.lints.rust]
 missing_docs = { level = "warn" }
diff --git a/llama-cpp-2/src/model.rs b/llama-cpp-2/src/model.rs
index 5f412c25..9f01ac24 100644
--- a/llama-cpp-2/src/model.rs
+++ b/llama-cpp-2/src/model.rs
@@ -90,6 +90,15 @@ impl LlamaModel {
         self.token_to_str_with_size(token, 32)
     }
 
+    /// Convert single token to bytes.
+    ///
+    /// # Errors
+    ///
+    /// See [`TokenToStringError`] for more information.
+    pub fn token_to_bytes(&self, token: LlamaToken) -> Result<Vec<u8>, TokenToStringError> {
+        self.token_to_bytes_with_size(token, 32)
+    }
+
     /// Convert a vector of tokens to a single string.
     ///
     /// # Errors
@@ -211,22 +220,45 @@ impl LlamaModel {
         token: LlamaToken,
         buffer_size: usize,
     ) -> Result<String, TokenToStringError> {
+        let bytes = self.token_to_bytes_with_size(token, buffer_size)?;
+        Ok(String::from_utf8(bytes)?)
+    }
+
+    /// Convert a token to bytes with a specified buffer size.
+    ///
+    /// Generally you should use [`LlamaModel::token_to_bytes`] instead as 8 bytes is enough for most words and
+    /// the extra bytes do not really matter.
+    ///
+    /// # Errors
+    ///
+    /// - if the token type is unknown
+    /// - the resultant token is larger than `buffer_size`.
+    ///
+    /// # Panics
+    ///
+    /// - if `buffer_size` does not fit into a [`c_int`].
+    /// - if the returned size from llama-cpp does not fit into a [`usize`]. (this should never happen)
+    pub fn token_to_bytes_with_size(
+        &self,
+        token: LlamaToken,
+        buffer_size: usize,
+    ) -> Result<Vec<u8>, TokenToStringError> {
         if token == self.token_nl() {
-            return Ok(String::from("\n"));
+            return Ok(String::from("\n").into_bytes());
         }
 
         match self.token_type(token) {
             LlamaTokenType::Normal | LlamaTokenType::UserDefined => {}
             LlamaTokenType::Control => {
                 if token == self.token_bos() || token == self.token_eos() {
-                    return Ok(String::new());
+                    return Ok(Vec::new());
                 }
             }
             LlamaTokenType::Unknown
             | LlamaTokenType::Undefined
             | LlamaTokenType::Byte
             | LlamaTokenType::Unused => {
-                return Ok(String::new());
+                return Ok(Vec::new());
             }
         }
 
@@ -246,7 +278,7 @@ impl LlamaModel {
                 let mut bytes = string.into_bytes();
                 let len = usize::try_from(size).expect("size is positive and fits into usize");
                 bytes.truncate(len);
-                Ok(String::from_utf8(bytes)?)
+                Ok(bytes)
             }
         }
     }
diff --git a/simple/Cargo.toml b/simple/Cargo.toml
index 4e23632a..66008226 100644
--- a/simple/Cargo.toml
+++ b/simple/Cargo.toml
@@ -10,6 +10,7 @@ llama-cpp-2 = { path = "../llama-cpp-2", version = "0.1.46" }
 hf-hub = { workspace = true }
 clap = { workspace = true , features = ["derive"] }
 anyhow = { workspace = true }
+encoding_rs = { workspace = true }
 
 [features]
 cublas = ["llama-cpp-2/cublas"]
diff --git a/simple/src/main.rs b/simple/src/main.rs
index 8f7451f7..98f863aa 100644
--- a/simple/src/main.rs
+++ b/simple/src/main.rs
@@ -240,6 +240,9 @@ either reduce n_len or increase n_ctx"
 
     let t_main_start = ggml_time_us();
 
+    // The `Decoder`
+    let mut decoder = encoding_rs::UTF_8.new_decoder();
+
     while n_cur <= n_len {
         // sample the next token
         {
@@ -256,7 +259,11 @@ either reduce n_len or increase n_ctx"
                 break;
             }
 
-            print!("{}", model.token_to_str(new_token_id)?);
+            let output_bytes = model.token_to_bytes(new_token_id)?;
+            // use `Decoder.decode_to_string()` to avoid the intermediate buffer
+            let mut output_string = String::with_capacity(32);
+            let _decode_result = decoder.decode_to_string(&output_bytes, &mut output_string, false);
+            print!("{}", output_string);
             std::io::stdout().flush()?;
 
             batch.clear();