diff --git a/router/src/lib.rs b/router/src/lib.rs
index d375eafb..efc4d3ae 100644
--- a/router/src/lib.rs
+++ b/router/src/lib.rs
@@ -56,12 +56,15 @@ pub(crate) struct GenerateParameters {
     #[schema(exclusive_minimum = 0, exclusive_maximum = 512, default = "20")]
     pub max_new_tokens: u32,
     #[serde(default)]
-    #[schema(default = "None", example = false)]
+    #[schema(default = "null", example = false)]
     pub return_full_text: Option<bool>,
     #[serde(default)]
     #[schema(inline, max_items = 4, example = json ! (["photographer"]))]
     pub stop: Vec<String>,
     #[serde(default)]
+    #[schema(default = "null", example = "null")]
+    pub truncate: Option<i32>,
+    #[serde(default)]
     #[schema(default = "false", example = true)]
     pub watermark: bool,
     #[serde(default)]
@@ -86,6 +89,7 @@ fn default_parameters() -> GenerateParameters {
         max_new_tokens: default_max_new_tokens(),
         return_full_text: None,
         stop: Vec::new(),
+        truncate: None,
         watermark: false,
         details: false,
         seed: None,
diff --git a/router/src/server.rs b/router/src/server.rs
index 2ce5699d..ef10b7b1 100644
--- a/router/src/server.rs
+++ b/router/src/server.rs
@@ -73,6 +73,7 @@ async fn health(infer: Extension<Infer>) -> Result<(), (StatusCode, Json<ErrorRe
                 max_new_tokens: 1,
                 return_full_text: None,
                 stop: Vec::new(),
+                truncate: None,
                 watermark: false,
                 details: false,
                 seed: None,
diff --git a/router/src/validation.rs b/router/src/validation.rs
index a280e652..1a247caa 100644
--- a/router/src/validation.rs
+++ b/router/src/validation.rs
@@ -225,7 +225,9 @@ fn validate(
 
     // Get the number of tokens in the input
     match tokenizer.encode(request.inputs.clone(), true) {
-        Ok(encoding) => {
+        Ok(mut encoding) => {
+            encoding.truncate()
+
             let input_length = encoding.len();
             let total_tokens = input_length + max_new_tokens as usize;