diff --git a/.github/scripts/allocator_bench_alpine.sh b/.github/scripts/allocator_bench_alpine.sh
new file mode 100644
index 0000000..cb6b3db
--- /dev/null
+++ b/.github/scripts/allocator_bench_alpine.sh
@@ -0,0 +1,64 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+TOKIO_FEATURE=""
+if [[ "${TOKIO_MULTI_THREAD:-false}" == "true" ]]; then
+  TOKIO_FEATURE="tokio-multi-thread"
+fi
+
+docker run --rm \
+  -v "$PWD:/work" \
+  -w /work \
+  -e TOKIO_FEATURE="$TOKIO_FEATURE" \
+  -e PLATFORM_LABEL="${PLATFORM_LABEL:-unknown}" \
+  rust:nightly-alpine sh -lc '
+    set -eu
+    apk add --no-cache build-base pkgconfig time
+
+    build_features() {
+      allocator="$1"
+      features=""
+      if [ "$allocator" != "system" ]; then
+        features="$allocator"
+      fi
+      if [ -n "$TOKIO_FEATURE" ]; then
+        if [ -n "$features" ]; then
+          features="$features,$TOKIO_FEATURE"
+        else
+          features="$TOKIO_FEATURE"
+        fi
+      fi
+      echo "$features"
+    }
+
+    : > /work/alpine-results.tsv
+    for allocator in system jemalloc mimalloc_v2 mimalloc_v3; do
+      features="$(build_features "$allocator")"
+      if [ -n "$features" ]; then
+        output="$(/usr/bin/time -v cargo run --release --locked --features "$features" 2>&1 >/dev/null)"
+      else
+        output="$(/usr/bin/time -v cargo run --release --locked 2>&1 >/dev/null)"
+      fi
+      peak="$(echo "$output" | awk -F': ' '/Maximum resident set size/ {print $2; exit}')"
+      if [ -z "$peak" ]; then
+        echo "Failed to parse peak memory for $allocator" >&2
+        exit 1
+      fi
+      printf "%s\t%s\n" "$allocator" "$peak" >> /work/alpine-results.tsv
+    done
+  '
+
+{
+  echo "### ${PLATFORM_LABEL:-unknown} (tokio-multi-thread=${TOKIO_MULTI_THREAD:-false})"
+  echo ""
+  echo "| Allocator | Peak KB |"
+  echo "| --- | ---: |"
+  sort -n -k2,2 alpine-results.tsv | while IFS=$'\t' read -r allocator peak; do
+    echo "| $allocator | $peak |"
+  done
+  best="$(sort -n -k2,2 alpine-results.tsv | head -n1)"
+  best_allocator="${best%%$'\t'*}"
+  best_peak="${best#*$'\t'}"
+  echo ""
+  echo "**Best:** $best_allocator ($best_peak KB)"
+} >> "$GITHUB_STEP_SUMMARY"
diff --git a/.github/scripts/allocator_bench_unix.sh b/.github/scripts/allocator_bench_unix.sh
new file mode 100644
index 0000000..b91bc6c
--- /dev/null
+++ b/.github/scripts/allocator_bench_unix.sh
@@ -0,0 +1,88 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+TOKIO_FEATURE=""
+if [[ "${TOKIO_MULTI_THREAD:-false}" == "true" ]]; then
+  TOKIO_FEATURE="tokio-multi-thread"
+fi
+
+allocators=(system jemalloc mimalloc_v2 mimalloc_v3)
+
+if [[ "$RUNNER_OS" == "Linux" ]]; then
+  time_cmd=(/usr/bin/time -v)
+  parse_peak() {
+    awk -F': ' '/Maximum resident set size/ {print $2; exit}'
+  }
+else
+  time_cmd=(/usr/bin/time -l)
+  parse_peak() {
+    awk '/maximum resident set size/ {print $1; exit}'
+  }
+fi
+
+build_features() {
+  local allocator="$1"
+  local features=""
+  if [[ "$allocator" != "system" ]]; then
+    features="$allocator"
+  fi
+  if [[ -n "$TOKIO_FEATURE" ]]; then
+    if [[ -n "$features" ]]; then
+      features="$features,$TOKIO_FEATURE"
+    else
+      features="$TOKIO_FEATURE"
+    fi
+  fi
+  echo "$features"
+}
+
+run_one() {
+  local allocator="$1"
+  local features
+  features="$(build_features "$allocator")"
+
+  local feature_args=()
+  if [[ -n "$features" ]]; then
+    feature_args=(--features "$features")
+  fi
+
+  local output
+  output="$(${time_cmd[@]} cargo run --release --locked "${feature_args[@]}" 2>&1 >/dev/null)"
+
+  local peak
+  peak="$(echo "$output" | parse_peak)"
+  if [[ -z "$peak" ]]; then
+    echo "Failed to parse peak memory for $allocator" >&2
+    exit 1
+  fi
+
+  if [[ "$RUNNER_OS" != "Linux" ]]; then
+    peak=$((peak / 1024))
+  fi
+
+  printf "%s\t%s\n" "$allocator" "$peak" >> results.tsv
+}
+
+if [[ "$RUNNER_OS" == "Windows" ]]; then
+  allocators=(system mimalloc_v2 mimalloc_v3)
+fi
+
+: > results.tsv
+for allocator in "${allocators[@]}"; do
+  run_one "$allocator"
+done
+
+{
+  echo "### ${PLATFORM_LABEL:-unknown} (tokio-multi-thread=${TOKIO_MULTI_THREAD:-false})"
+  echo ""
+  echo "| Allocator | Peak KB |"
+  echo "| --- | ---: |"
+  sort -n -k2,2 results.tsv | while IFS=$'\t' read -r allocator peak; do
+    echo "| $allocator | $peak |"
+  done
+  best="$(sort -n -k2,2 results.tsv | head -n1)"
+  best_allocator="${best%%$'\t'*}"
+  best_peak="${best#*$'\t'}"
+  echo ""
+  echo "**Best:** $best_allocator ($best_peak KB)"
+} >> "$GITHUB_STEP_SUMMARY"
diff --git a/.github/scripts/allocator_bench_windows.ps1 b/.github/scripts/allocator_bench_windows.ps1
new file mode 100644
index 0000000..1939058
--- /dev/null
+++ b/.github/scripts/allocator_bench_windows.ps1
@@ -0,0 +1,57 @@
+$tokioOn = $env:TOKIO_MULTI_THREAD
+$tokioFeature = if ($tokioOn -eq "true") { "tokio-multi-thread" } else { "" }
+
+$allocators = @("system", "mimalloc_v2", "mimalloc_v3")
+
+function Build-Features([string]$allocator, [string]$tokio) {
+  $features = @()
+  if ($allocator -ne "system") { $features += $allocator }
+  if ($tokio) { $features += $tokio }
+  return ($features -join ",")
+}
+
+function Run-One([string]$allocator, [string]$features) {
+  $args = @("run", "--release", "--locked")
+  if ($features) { $args += "--features"; $args += $features }
+
+  $peak = 0
+  $proc = Start-Process -FilePath "cargo" -ArgumentList $args -PassThru -NoNewWindow
+  while (-not $proc.HasExited) {
+    Start-Sleep -Milliseconds 200
+    try {
+      $p = Get-Process -Id $proc.Id -ErrorAction Stop
+      $current = [math]::Max($p.WorkingSet64, $p.PeakWorkingSet64)
+      if ($current -gt $peak) { $peak = $current }
+    } catch { }
+  }
+  $peakKb = [math]::Floor($peak / 1kb)
+  Add-Content -Path results.tsv -Value "$allocator`t$peakKb"
+}
+
+if (Test-Path results.tsv) { Remove-Item results.tsv -Force }
+foreach ($allocator in $allocators) {
+  $features = Build-Features $allocator $tokioFeature
+  Run-One $allocator $features
+}
+
+$rows = Get-Content results.tsv | ForEach-Object {
+  $parts = $_ -split "`t"
+  [pscustomobject]@{
+    Allocator = $parts[0]
+    PeakKB = [int]$parts[1]
+  }
+}
+$sorted = $rows | Sort-Object PeakKB
+$best = $sorted | Select-Object -First 1
+
+$summary = @()
+$summary += "### $($env:PLATFORM_LABEL) (tokio-multi-thread=$tokioOn)"
+$summary += ""
+$summary += "| Allocator | Peak KB |"
+$summary += "| --- | ---: |"
+foreach ($row in $sorted) {
+  $summary += "| $($row.Allocator) | $($row.PeakKB) |"
+}
+$summary += ""
+$summary += "**Best:** $($best.Allocator) ($($best.PeakKB) KB)"
+$summary -join "`n" | Add-Content $env:GITHUB_STEP_SUMMARY
diff --git a/.github/workflows/allocator-bench.yml b/.github/workflows/allocator-bench.yml
new file mode 100644
index 0000000..7048794
--- /dev/null
+++ b/.github/workflows/allocator-bench.yml
@@ -0,0 +1,73 @@
+name: Allocator Bench
+on:
+  workflow_dispatch:
+concurrency:
+  group: ${{ github.workflow }}-${{ github.run_id }}
+  cancel-in-progress: true
+permissions:
+  contents: read
+jobs:
+  bench:
+    name: ${{ matrix.name }} (tokio-multi-thread=${{ matrix.tokio_multi_thread }})
+    runs-on: ${{ matrix.runner }}
+    strategy:
+      fail-fast: false
+      matrix:
+        tokio_multi_thread: [true, false]
+        include:
+          - name: ubuntu-24.04-arm-1
+            platform: ubuntu-24.04-arm
+            runner: ubuntu-24.04-arm
+            alpine: false
+          - name: ubuntu-24.04-arm-2
+            platform: ubuntu-24.04-arm
+            runner: ubuntu-24.04-arm
+            alpine: false
+          - name: windows-2025
+            platform: windows-2025
+            runner: windows-2025
+            alpine: false
+          - name: windows-11-arm
+            platform: windows-11-arm
+            runner: windows-11-arm
+            alpine: false
+          - name: macos-15-x86_64
+            platform: macos-15 x86_64
+            runner: macos-15-large
+            alpine: false
+          - name: macos-15-arm64
+            platform: macos-15 arm64
+            runner: macos-15
+            alpine: false
+          - name: alpine-dockerized
+            platform: alpine (dockerized)
+            runner: ubuntu-24.04
+            alpine: true
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+      - if: ${{ !matrix.alpine }}
+        uses: actions-rust-lang/setup-rust-toolchain@v1
+        with:
+          toolchain: beta
+          cache: true
+          cache-key: allocator-bench-${{ matrix.runner }}
+      - if: ${{ !matrix.alpine && runner.os != 'Windows' }}
+        shell: bash
+        env:
+          TOKIO_MULTI_THREAD: ${{ matrix.tokio_multi_thread }}
+          PLATFORM_LABEL: ${{ matrix.platform }}
+        run: bash .github/scripts/allocator_bench_unix.sh
+      - if: ${{ !matrix.alpine && runner.os == 'Windows' }}
+        shell: pwsh
+        env:
+          TOKIO_MULTI_THREAD: ${{ matrix.tokio_multi_thread }}
+          PLATFORM_LABEL: ${{ matrix.platform }}
+        run: ./.github/scripts/allocator_bench_windows.ps1
+      - if: ${{ matrix.alpine }}
+        shell: bash
+        env:
+          TOKIO_MULTI_THREAD: ${{ matrix.tokio_multi_thread }}
+          PLATFORM_LABEL: ${{ matrix.platform }}
+        run: bash .github/scripts/allocator_bench_alpine.sh
\ No newline at end of file