diff --git a/sjsonnet/src/sjsonnet/Format.scala b/sjsonnet/src/sjsonnet/Format.scala
index 77377dac..c8a30990 100644
--- a/sjsonnet/src/sjsonnet/Format.scala
+++ b/sjsonnet/src/sjsonnet/Format.scala
@@ -286,7 +286,13 @@ object Format {
       else if (signedConversion && formatted.signCharacter) "+" + lhs
       else lhs
 
-    val missingWidth = formatted.widthOr(-1) - lhs2.length - mhs.length - rhs.length
+    val missingWidth =
+      if (numeric) formatted.widthOr(-1) - lhs2.length - mhs.length - rhs.length
+      else
+        formatted.widthOr(-1) -
+        lhs2.codePointCount(0, lhs2.length) -
+        mhs.codePointCount(0, mhs.length) -
+        rhs.codePointCount(0, rhs.length)
 
     if (missingWidth <= 0) {
       // Avoid unnecessary string concatenation when parts are empty
diff --git a/sjsonnet/test/resources/new_test_suite/format_width_codepoint.jsonnet b/sjsonnet/test/resources/new_test_suite/format_width_codepoint.jsonnet
new file mode 100644
index 00000000..bb469103
--- /dev/null
+++ b/sjsonnet/test/resources/new_test_suite/format_width_codepoint.jsonnet
@@ -0,0 +1,11 @@
+// Width padding must count codepoints, not UTF-16 code units.
+// Supplementary characters (surrogate pairs) count as 1 codepoint.
+[
+  std.assertEqual("%5s" % "😀", "    😀"),
+  std.assertEqual("%-5s" % "😀", "😀    "),
+  std.assertEqual("%5s" % "a😀b", "  a😀b"),
+  std.assertEqual("%3s" % "😀😀", " 😀😀"),
+  std.assertEqual("%5.1s" % "😀x", "    😀"),
+  std.assertEqual("%5s" % "hello", "hello"),
+  std.assertEqual("%10s" % "hello", "     hello"),
+]
diff --git a/sjsonnet/test/resources/new_test_suite/format_width_codepoint.jsonnet.golden b/sjsonnet/test/resources/new_test_suite/format_width_codepoint.jsonnet.golden
new file mode 100644
index 00000000..15f719b5
--- /dev/null
+++ b/sjsonnet/test/resources/new_test_suite/format_width_codepoint.jsonnet.golden
@@ -0,0 +1,9 @@
+[
+   true,
+   true,
+   true,
+   true,
+   true,
+   true,
+   true
+]