diff --git a/src/uu/numfmt/src/format.rs b/src/uu/numfmt/src/format.rs
index 9c03801ef7..deb7a045ff 100644
--- a/src/uu/numfmt/src/format.rs
+++ b/src/uu/numfmt/src/format.rs
@@ -57,7 +57,11 @@ fn find_valid_number_with_suffix(s: &str, unit: Unit) -> Option<&str> {
     let accepts_suffix = unit != Unit::None;
     let accepts_i = [Unit::Auto, Unit::Iec(true)].contains(&unit);
 
-    let mut characters = s.chars().skip(numeric_part.len());
+    // Look at the characters immediately after the numeric part. `numeric_part` is a prefix of
+    // `s`, so its byte length is a valid char boundary to slice/iterate from. (Using it as a
+    // *character* count for `chars().skip(..)` was wrong for multi-byte numeric parts, e.g. a
+    // multi-byte decimal separator, and led to slicing inside a multi-byte suffix char. See #13937.)
+    let mut characters = s[numeric_part.len()..].chars();
     let potential_suffix = characters.next();
     let potential_i = characters.next();
 
@@ -66,14 +70,11 @@ fn find_valid_number_with_suffix(s: &str, unit: Unit) -> Option<&str> {
     }
 
     match (potential_suffix, potential_i) {
-        (Some(suffix), None) if RawSuffix::try_from(&suffix).is_ok() => {
-            Some(&s[..=numeric_part.len()])
-        }
         (Some(suffix), Some('i')) if accepts_i && RawSuffix::try_from(&suffix).is_ok() => {
-            Some(&s[..numeric_part.len() + 2])
+            Some(&s[..numeric_part.len() + suffix.len_utf8() + 'i'.len_utf8()])
         }
-        (Some(suffix), Some(_)) if RawSuffix::try_from(&suffix).is_ok() => {
-            Some(&s[..=numeric_part.len()])
+        (Some(suffix), _) if RawSuffix::try_from(&suffix).is_ok() => {
+            Some(&s[..numeric_part.len() + suffix.len_utf8()])
         }
         _ => Some(numeric_part),
     }
diff --git a/tests/by-util/test_numfmt.rs b/tests/by-util/test_numfmt.rs
index 4db8694477..1b1fd6ad3f 100644
--- a/tests/by-util/test_numfmt.rs
+++ b/tests/by-util/test_numfmt.rs
@@ -1614,6 +1614,20 @@ fn test_locale_fr_rejects_period() {
         .stderr_contains("invalid");
 }
 
+#[test]
+#[cfg_attr(wasi_runner, ignore = "WASI: locale env vars not propagated")]
+fn test_locale_multibyte_separator_invalid_suffix_does_not_panic() {
+    // Regression test for #13937: under a locale with a multi-byte decimal separator (Arabic
+    // '٫', U+066B, 2 bytes), an input like "1٫€K" made numfmt slice inside a multi-byte
+    // character while parsing the suffix and panic. It must instead reject the malformed suffix
+    // like GNU (non-zero exit, "invalid suffix in input"), without crashing.
+    new_ucmd!()
+        .env("LC_ALL", "ar_SA.UTF-8")
+        .args(&["--from=auto", "1٫€K"])
+        .fails()
+        .stderr_contains("invalid suffix in input");
+}
+
 #[test]
 fn test_locale_c_uses_period() {
     // C locale should still use '.' as usual