diff --git a/src/utils/types.cpp b/src/utils/types.cpp
index 4163c21..b5710e4 100644
--- a/src/utils/types.cpp
+++ b/src/utils/types.cpp
@@ -52,8 +52,21 @@ fp16_t _f32_to_f16(float val) {
         // Infinity
         return fp16_t{static_cast<uint16_t>(sign | 0x7C00)};
     } else if (exponent >= -14) { // Normalized case
-        return fp16_t{(uint16_t)(sign | ((exponent + 15) << 10) | (mantissa >> 13))};
-    } else if (exponent >= -24) {
+    // --- START MODIFICATION ---
+    // Add 0x1000 (2^12), which is half of the value of the 13th bit.
+    // This effectively rounds the 10-bit mantissa to the nearest value.
+    uint32_t rounded_mantissa = mantissa + 0x1000;
+
+    // Check for overflow in the mantissa after rounding
+    if (rounded_mantissa & 0x800000) {
+        // If mantissa overflows, we need to increment the exponent
+        // and reset mantissa. This is rare but important for correctness.
+        return fp16_t{(uint16_t)(sign | ((exponent + 15 + 1) << 10))};
+    }
+    
+    return fp16_t{(uint16_t)(sign | ((exponent + 15) << 10) | (rounded_mantissa >> 13))};
+    // --- END MODIFICATION ---
+} else if (exponent >= -24) {
         mantissa |= 0x800000; // Add implicit leading 1
         mantissa >>= (-14 - exponent);
         return fp16_t{(uint16_t)(sign | (mantissa >> 13))};