diff --git a/Cargo.lock b/Cargo.lock
index 2d6665f8f..7a1b242f4 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -126,6 +126,8 @@ name = "fish-printf"
 version = "0.2.1"
 dependencies = [
  "libc",
+ "unicode-segmentation",
+ "unicode-width",
  "widestring",
 ]
 
@@ -546,6 +548,18 @@ version = "1.0.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "adb9e6ca4f869e1180728b7950e35922a7fc6397f7b641499e8f3ef06e50dc83"
 
+[[package]]
+name = "unicode-segmentation"
+version = "1.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f6ccf251212114b54433ec949fd6a7841275f9ada20dddd2f29e9ceea4501493"
+
+[[package]]
+name = "unicode-width"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fc81956842c57dac11422a97c3b8195a1ff727f06e85c84ed2e8aa277c9a0fd"
+
 [[package]]
 name = "unix_path"
 version = "1.0.1"
diff --git a/printf/Cargo.toml b/printf/Cargo.toml
index 09a6413ca..cddbb956b 100644
--- a/printf/Cargo.toml
+++ b/printf/Cargo.toml
@@ -10,6 +10,8 @@ license = "MIT"
 [dependencies]
 libc = "0.2.155"
 widestring = { version = "1.0.2", optional = true }
+unicode-segmentation = "1.12.0"
+unicode-width = "0.2.0"
 
 [lints]
 workspace = true
diff --git a/printf/src/lib.rs b/printf/src/lib.rs
index 733d7a193..b7d4ac5cb 100644
--- a/printf/src/lib.rs
+++ b/printf/src/lib.rs
@@ -71,7 +71,7 @@ macro_rules! sprintf {
 /// - `args`: Iterator over the arguments to format.
 ///
 /// # Returns
-/// A `Result` which is `Ok` containing the number of characters written on success, or an `Error`.
+/// A `Result` which is `Ok` containing the width of the string written on success, or an `Error`.
 ///
 /// # Example
 ///
diff --git a/printf/src/printf_impl.rs b/printf/src/printf_impl.rs
index 8a09137ac..1f26e9264 100644
--- a/printf/src/printf_impl.rs
+++ b/printf/src/printf_impl.rs
@@ -5,6 +5,8 @@
 use std::fmt::{self, Write};
 use std::mem;
 use std::result::Result;
+use unicode_segmentation::UnicodeSegmentation;
+use unicode_width::UnicodeWidthStr;
 
 #[cfg(feature = "widestring")]
 use widestring::Utf32Str as wstr;
@@ -382,7 +384,7 @@ pub fn sprintf_locale(
         }
 
         // Read field width. We do not support $.
-        let width = if s.at(0) == Some('*') {
+        let desired_width = if s.at(0) == Some('*') {
             let arg_width = args.next().ok_or(Error::MissingArg)?.as_sint()?;
             s.advance_by(1);
             if arg_width < 0 {
@@ -397,7 +399,7 @@ pub fn sprintf_locale(
         };
 
         // Optionally read precision. We do not support $.
-        let mut prec: Option<usize> = if s.at(0) == Some('.') && s.at(1) == Some('*') {
+        let mut desired_precision: Option<usize> = if s.at(0) == Some('.') && s.at(1) == Some('*') {
             // "A negative precision is treated as though it were missing."
             // Here we assume the precision is always signed.
             s.advance_by(2);
@@ -410,7 +412,7 @@ pub fn sprintf_locale(
             None
         };
         // Disallow precisions larger than i32::MAX, in keeping with C.
-        if prec.unwrap_or(0) > i32::MAX as usize {
+        if desired_precision.unwrap_or(0) > i32::MAX as usize {
             return Err(Error::Overflow);
         }
 
@@ -429,7 +431,7 @@ pub fn sprintf_locale(
         // "If a precision is given with a numeric conversion (d, i, o, u, i, x, and X),
         // the 0 flag is ignored." p is included here.
         let spec_is_numeric = matches!(conv_spec, CS::d | CS::u | CS::o | CS::p | CS::x | CS::X);
-        if spec_is_numeric && prec.is_some() {
+        if spec_is_numeric && desired_precision.is_some() {
             flags.zero_pad = false;
         }
 
@@ -443,13 +445,22 @@ pub fn sprintf_locale(
             CS::e | CS::f | CS::g | CS::a | CS::E | CS::F | CS::G | CS::A => {
                 // Floating point types handle output on their own.
                 let float = arg.as_float()?;
-                let len = format_float(f, float, width, prec, flags, locale, conv_spec, buf)?;
+                let len = format_float(
+                    f,
+                    float,
+                    desired_width,
+                    desired_precision,
+                    flags,
+                    locale,
+                    conv_spec,
+                    buf,
+                )?;
                 out_len = out_len.checked_add(len).ok_or(Error::Overflow)?;
                 continue 'main;
             }
             CS::p => {
                 const PTR_HEX_DIGITS: usize = 2 * mem::size_of::<*const u8>();
-                prec = prec.map(|p| p.max(PTR_HEX_DIGITS));
+                desired_precision = desired_precision.map(|p| p.max(PTR_HEX_DIGITS));
                 let uint = arg.as_uint()?;
                 if uint != 0 {
                     prefix = "0x";
@@ -479,8 +490,8 @@ pub fn sprintf_locale(
                 if uint != 0 {
                     write!(buf, "{:o}", uint)?;
                 }
-                if flags.alt_form && prec.unwrap_or(0) <= buf.len() + 1 {
-                    prec = Some(buf.len() + 1);
+                if flags.alt_form && desired_precision.unwrap_or(0) <= buf.len() + 1 {
+                    desired_precision = Some(buf.len() + 1);
                 }
                 buf
             }
@@ -514,10 +525,38 @@ pub fn sprintf_locale(
             CS::s => {
                 // also 'S'
                 let s = arg.as_str(buf)?;
-                let p = prec.unwrap_or(s.len()).min(s.len());
-                prec = Some(p);
                 flags.zero_pad = false;
-                &s[..p]
+                match desired_precision {
+                    Some(precision) => {
+                        // from man printf(3)
+                        // "the maximum number of characters to be printed from a string"
+                        // We interpret this to mean the maximum width when printed, as defined by
+                        // Unicode grapheme cluster width.
+                        let mut byte_len = 0;
+                        let mut width = 0;
+                        let mut graphemes = s.graphemes(true);
+                        // Iteratively add single grapheme clusters as long as the fit within the
+                        // width limited by precision.
+                        while width < precision {
+                            match graphemes.next() {
+                                Some(grapheme) => {
+                                    let grapheme_width = grapheme.width();
+                                    if width + grapheme_width <= precision {
+                                        byte_len += grapheme.len();
+                                        width += grapheme_width;
+                                    } else {
+                                        break;
+                                    }
+                                }
+                                None => break,
+                            }
+                        }
+                        let p = precision.min(width);
+                        desired_precision = Some(p);
+                        &s[..byte_len]
+                    }
+                    None => s,
+                }
             }
         };
         // Numeric output should be empty iff the value is 0.
@@ -528,23 +567,26 @@ pub fn sprintf_locale(
         // Decide if we want to apply thousands grouping to the body, and compute its size.
         // Note we have already errored out if grouped is set and this is non-numeric.
         let wants_grouping = flags.grouped && locale.thousands_sep.is_some();
-        let body_len = match wants_grouping {
+        let body_width = match wants_grouping {
+            // We assume that text representing numbers is ASCII, so len == width.
             true => body.len() + locale.separator_count(body.len()),
-            false => body.len(),
+            false => body.width(),
         };
 
         // Resolve the precision.
         // In the case of a non-numeric conversion, update the precision to at least the
         // length of the string.
-        let prec = if !spec_is_numeric {
-            prec.unwrap_or(body_len)
+        let desired_precision = if !spec_is_numeric {
+            desired_precision.unwrap_or(body_width)
         } else {
-            prec.unwrap_or(1).max(body_len)
+            desired_precision.unwrap_or(1).max(body_width)
         };
 
-        let prefix_len = prefix.len();
-        let unpadded_width = prefix_len.checked_add(prec).ok_or(Error::Overflow)?;
-        let width = width.max(unpadded_width);
+        let prefix_width = prefix.width();
+        let unpadded_width = prefix_width
+            .checked_add(desired_precision)
+            .ok_or(Error::Overflow)?;
+        let width = desired_width.max(unpadded_width);
 
         // Pad on the left with spaces to the desired width?
         if !flags.left_adj && !flags.zero_pad {
@@ -560,7 +602,8 @@ pub fn sprintf_locale(
         }
 
         // Pad on the left to the given precision?
-        pad(f, '0', prec, body_len)?;
+        // TODO: why pad with 0 here?
+        pad(f, '0', desired_precision, body_width)?;
 
         // Output the actual value, perhaps with grouping.
         if wants_grouping {
diff --git a/printf/src/tests.rs b/printf/src/tests.rs
index 85b966af5..036fe0d3f 100644
--- a/printf/src/tests.rs
+++ b/printf/src/tests.rs
@@ -13,6 +13,7 @@ macro_rules! sprintf_check {
         $(,)? // optional trailing comma
     ) => {
         {
+            use unicode_width::UnicodeWidthStr;
             let mut target = String::new();
             let mut args = [$($arg.to_arg()),*];
             let len = $crate::printf_c_locale(
@@ -20,7 +21,7 @@ macro_rules! sprintf_check {
                 $fmt.as_ref() as &str,
                 &mut args,
             ).expect("printf failed");
-            assert!(len == target.len(), "Wrong length returned: {} vs {}", len, target.len());
+            assert_eq!(len, target.width(), "Wrong length returned");
             target
         }
     };
@@ -735,6 +736,18 @@ fn test_huge_precision_g() {
     sprintf_err!("%.2147483648g", f => Error::Overflow);
 }
 
+#[test]
+fn test_non_ascii() {
+    assert_fmt!("%3s", "ö" => "  ö");
+    assert_fmt!("%3s", "🇺🇳" => " 🇺🇳");
+    assert_fmt!("%.3s", "🇺🇳🇺🇳" => "🇺🇳");
+    assert_fmt!("%.3s", "a🇺🇳" => "a🇺🇳");
+    assert_fmt!("%.3s", "aa🇺🇳" => "aa");
+    assert_fmt!("%3.3s", "aa🇺🇳" => " aa");
+    assert_fmt!("%.1s", "𒈙a" => "𒈙");
+    assert_fmt!("%3.3s", "👨‍👨‍👧‍👧" => " 👨‍👨‍👧‍👧");
+}
+
 #[test]
 fn test_errors() {
     use Error::*;
diff --git a/tests/checks/printf.fish b/tests/checks/printf.fish
index c905a4ebe..99c7a3fd0 100644
--- a/tests/checks/printf.fish
+++ b/tests/checks/printf.fish
@@ -154,3 +154,21 @@ printf '%b\n' '\0057foo\0057bar\0057'
 
 printf %18446744073709551616s
 # CHECKERR: Number out of range
+
+# Test non-ASCII behavior
+printf '|%3s|\n' 'ö'
+# CHECK: |  ö|
+printf '|%3s|\n' '🇺🇳'
+#CHECK: | 🇺🇳|
+printf '|%.3s|\n' '🇺🇳🇺🇳'
+#CHECK: |🇺🇳|
+printf '|%.3s|\n' 'a🇺🇳'
+#CHECK: |a🇺🇳|
+printf '|%.3s|\n' 'aa🇺🇳'
+#CHECK: |aa|
+printf '|%3.3s|\n' 'aa🇺🇳'
+#CHECK: | aa|
+printf '|%.1s|\n' '𒈙a'
+#CHECK: |𒈙|
+printf '|%3.3s|\n' '👨‍👨‍👧‍👧'
+#CHECK: | 👨‍👨‍👧‍👧|