diff --git a/CHANGELOG.md b/CHANGELOG.md
index f02f3be..7bad628 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,7 @@
+# 2.4.3
+
+- Optimize Argon2 on the Web/JS platform by replacing `BigInt` arithmetic in the compression function with 32-bit integer math
+
 # 2.4.2
 
 - Bump `convertlib` to `^3.5.1`
diff --git a/lib/src/algorithms/argon2/argon2_32bit.dart b/lib/src/algorithms/argon2/argon2_32bit.dart
index 4a197b9..dac272e 100644
--- a/lib/src/algorithms/argon2/argon2_32bit.dart
+++ b/lib/src/algorithms/argon2/argon2_32bit.dart
@@ -421,13 +421,36 @@ class Argon2Internal {
   }
 
   /// `v[x] += v[y] + 2 * ((v[x] & _mask32) * (v[y] & _mask32))`
+  ///
+  /// The 64-bit words are stored as (low, high) 32-bit pairs. To stay correct
+  /// on the web where integers are IEEE-754 doubles, the low-word product is
+  /// built from 16-bit limbs and every carry is propagated with `~/`/`%` by
+  /// `2^32`; all intermediate values remain below `2^53`, so no precision is
+  /// lost and no `BigInt` allocation is needed.
   static void _fBlaMka(Uint32List v, int x, int y) {
-    var t = (BigInt.from(v[x]) * BigInt.from(v[y])) << 1;
-    t += (BigInt.from(v[x + 1]) << 32) + BigInt.from(v[x]);
-    t += (BigInt.from(v[y + 1]) << 32) + BigInt.from(v[y]);
-
-    v[x] = t.toUnsigned(32).toInt();
-    v[x + 1] = (t >> 32).toUnsigned(32).toInt();
+    int xl = v[x], xh = v[x + 1];
+    int yl = v[y], yh = v[y + 1];
+
+    // Full 32x32 -> 64 product of the low words via 16-bit limbs.
+    int a0 = xl & _mask16, a1 = xl >>> 16;
+    int b0 = yl & _mask16, b1 = yl >>> 16;
+    int m0 = a0 * b0;
+    int m1 = a0 * b1 + a1 * b0;
+    int m2 = a1 * b1;
+    int lo = m0 + (m1 % 0x10000) * 0x10000;
+    int pLo = lo % 0x100000000;
+    int pHi = m2 + m1 ~/ 0x10000 + lo ~/ 0x100000000;
+
+    // Double the product: 2 * p (mod 2^64).
+    int dLo = pLo * 2;
+    int dHi = pHi * 2 + dLo ~/ 0x100000000;
+    dLo %= 0x100000000;
+
+    // v[x] = (xl,xh) + (yl,yh) + (dLo,dHi) (mod 2^64).
+    int sumLo = xl + yl + dLo;
+    int sumHi = xh + yh + dHi + sumLo ~/ 0x100000000;
+    v[x] = sumLo; // Uint32List store truncates to the low 32 bits
+    v[x + 1] = sumHi;
   }
 
   // v[k] = (v[i] << (64 - n)) | (v[i] >>> n)
diff --git a/pubspec.yaml b/pubspec.yaml
index a08966d..08aa35d 100644
--- a/pubspec.yaml
+++ b/pubspec.yaml
@@ -1,7 +1,7 @@
 name: hashlib
 description: Secure hash functions, checksum generators, and key derivation algorithms optimized for Dart.
 homepage: https://github.com/bitanon/hashlib
-version: 2.4.2
+version: 2.4.3
 
 environment:
   sdk: '>=2.19.0 <4.0.0'