diff --git a/bench/src/main/scala/perfio/BufferedInputNumBenchmark.scala b/bench/src/main/scala/perfio/BufferedInputNumBenchmark.scala
index f472942..c7ecea6 100644
--- a/bench/src/main/scala/perfio/BufferedInputNumBenchmark.scala
+++ b/bench/src/main/scala/perfio/BufferedInputNumBenchmark.scala
@@ -1,5 +1,7 @@
 package perfio
 
+import com.esotericsoftware.kryo.io.{ByteBufferInput, Input}
+import com.esotericsoftware.kryo.unsafe.UnsafeByteBufferInput
 import org.openjdk.jmh.annotations.*
 import org.openjdk.jmh.infra.*
 
@@ -8,7 +10,10 @@ import java.nio.{ByteBuffer, ByteOrder}
 import java.util.concurrent.TimeUnit
 
 @BenchmarkMode(Array(Mode.AverageTime))
-@Fork(value = 1, jvmArgsAppend = Array("-Xmx12g", "-Xss32M", "-XX:+UnlockExperimentalVMOptions", "-XX:+UseZGC", "--enable-native-access=ALL-UNNAMED", "--add-modules", "jdk.incubator.vector"))
+@Fork(value = 1, jvmArgsAppend = Array("-Xmx12g", "-Xss32M", "-XX:+UnlockExperimentalVMOptions", "-XX:+UseZGC",
+  "--enable-native-access=ALL-UNNAMED", "--add-modules", "jdk.incubator.vector",
+  "--add-opens", "java.base/sun.nio.ch=ALL-UNNAMED" // for KryoUnsafe
+))
 @Threads(1)
 @Warmup(iterations = 7, time = 1)
 @Measurement(iterations = 7, time = 1)
@@ -47,6 +52,7 @@ class BufferedInputNumBenchmark extends BenchUtil {
     bin.close()
   }
 
+  // JDK: always big endian
   def run(bh: Blackhole, din: DataInputStream): Unit = {
     var i = 0
     while(i < count) {
@@ -58,6 +64,18 @@ class BufferedInputNumBenchmark extends BenchUtil {
     din.close()
   }
 
+  // Kryo: little endian (safe) or native endian (unsafe)
+  def run(bh: Blackhole, din: Input): Unit = {
+    var i = 0
+    while(i < count) {
+      bh.consume(din.readByte())
+      bh.consume(din.readInt())
+      bh.consume(din.readLong())
+      i += 1
+    }
+    din.close()
+  }
+
   def run(bh: Blackhole, buf: ByteBuffer): Unit = {
     buf.order(ByteOrder.BIG_ENDIAN)
     var i = 0
@@ -72,6 +90,10 @@ class BufferedInputNumBenchmark extends BenchUtil {
   @Benchmark
   def array_DataInputStream(bh: Blackhole): Unit = run(bh, new DataInputStream(new ByteArrayInputStream(testData)))
   @Benchmark
+  def array_Kryo(bh: Blackhole): Unit = run(bh, new ByteBufferInput(testData))
+  @Benchmark
+  def array_KryoUnsafe(bh: Blackhole): Unit = run(bh, new UnsafeByteBufferInput(testData))
+  @Benchmark
   def array_ByteBuffer(bh: Blackhole): Unit = run(bh, ByteBuffer.wrap(testData))
   @Benchmark
   def array_BufferedInput(bh: Blackhole): Unit = run(bh, BufferedInput.of(new ByteArrayInputStream(testData)))
diff --git a/bench/src/main/scala/perfio/BufferedOutputNumBenchmark.scala b/bench/src/main/scala/perfio/BufferedOutputNumBenchmark.scala
index df6cab7..0db9f7e 100644
--- a/bench/src/main/scala/perfio/BufferedOutputNumBenchmark.scala
+++ b/bench/src/main/scala/perfio/BufferedOutputNumBenchmark.scala
@@ -1,5 +1,7 @@
 package perfio
 
+import com.esotericsoftware.kryo.io.{ByteBufferOutput, Output}
+import com.esotericsoftware.kryo.unsafe.UnsafeByteBufferOutput
 import org.openjdk.jmh.annotations.*
 import org.openjdk.jmh.infra.*
 import perfio.internal.BufferUtil
@@ -10,7 +12,10 @@ import java.nio.file.Paths
 import java.util.concurrent.TimeUnit
 
 @BenchmarkMode(Array(Mode.AverageTime))
-@Fork(value = 1, jvmArgsAppend = Array("-Xmx12g", "-Xss32M", "-XX:+UnlockExperimentalVMOptions", "-XX:+UseZGC", "--enable-native-access=ALL-UNNAMED", "--add-modules", "jdk.incubator.vector"))
+@Fork(value = 1, jvmArgsAppend = Array("-Xmx12g", "-Xss32M", "-XX:+UnlockExperimentalVMOptions", "-XX:+UseZGC",
+  "--enable-native-access=ALL-UNNAMED", "--add-modules", "jdk.incubator.vector",
+  "--add-opens", "java.base/sun.nio.ch=ALL-UNNAMED" // for KryoUnsafe
+))
 @Threads(1)
 @Warmup(iterations = 15, time = 1)
 @Measurement(iterations = 15, time = 1)
@@ -21,6 +26,7 @@ class BufferedOutputNumBenchmark extends BenchUtil:
   val count = 20000000
   val byteSize = count * 13
 
+  // JDK: always big endian
   private def writeTo(out: DataOutputStream): Unit =
     var i = 0
     while i < count do
@@ -30,6 +36,16 @@ class BufferedOutputNumBenchmark extends BenchUtil:
       i += 1
     out.close()
 
+  // Kryo: little endian (safe) or native endian (unsafe)
+  private def writeTo(out: Output): Unit =
+    var i = 0
+    while i < count do
+      out.writeByte(i)
+      out.writeInt(i+100)
+      out.writeLong(i+101)
+      i += 1
+    out.close()
+
   private def writeTo(out: BufferedOutput): Unit =
     var i = 0
     while i < count do
@@ -65,6 +81,36 @@ class BufferedOutputNumBenchmark extends BenchUtil:
     bh.consume(bout.getSize)
     bh.consume(bout.getBuffer)
 
+  @Benchmark
+  def array_Kryo_growing(bh: Blackhole): Unit =
+    val bout = new MyByteArrayOutputStream
+    val out = new ByteBufferOutput(bout)
+    writeTo(out)
+    bh.consume(bout.getSize)
+    bh.consume(bout.getBuffer)
+
+  @Benchmark
+  def array_Kryo_preallocated(bh: Blackhole): Unit =
+    val bb = ByteBuffer.allocate(count * 13)
+    val out = new ByteBufferOutput(bb)
+    writeTo(out)
+    bh.consume(out.position())
+
+  @Benchmark
+  def array_KryoUnsafe_growing(bh: Blackhole): Unit =
+    val bout = new MyByteArrayOutputStream
+    val out = new UnsafeByteBufferOutput(bout)
+    writeTo(out)
+    bh.consume(bout.getSize)
+    bh.consume(bout.getBuffer)
+
+  @Benchmark
+  def array_KryoUnsafe_preallocated(bh: Blackhole): Unit =
+    val bb = ByteBuffer.allocate(count * 13)
+    val out = new UnsafeByteBufferOutput(count * 13)
+    writeTo(out)
+    bh.consume(out.position())
+
   @Benchmark
   def array_FlushingBufferedOutput_growing(bh: Blackhole): Unit =
     val bout = new MyByteArrayOutputStream
diff --git a/build.sbt b/build.sbt
index d71cb1d..7164e03 100644
--- a/build.sbt
+++ b/build.sbt
@@ -93,7 +93,8 @@ lazy val bench = (project in file("bench"))
   .settings(
     scalacOptions ++= Seq("-feature", "-opt:l:inline", "-opt-inline-from:perfio.*"),
     libraryDependencies ++= Seq(
-      "com.google.guava" % "guava" % "33.3.0-jre"
+      "com.google.guava" % "guava" % "33.3.0-jre",
+      "com.esotericsoftware" % "kryo" % "5.6.2",
     ),
     name := "perfio-bench",
     publish / skip := true,