diff --git a/docker/Dockerfile.arm64 b/docker/Dockerfile.arm64
new file mode 100644
index 0000000..fd50d92
--- /dev/null
+++ b/docker/Dockerfile.arm64
@@ -0,0 +1,51 @@
+# Cross-build parakeet.cpp for arm64 (aarch64) via docker buildx + qemu.
+#
+#   docker buildx build --platform linux/arm64 -f docker/Dockerfile.arm64 \
+#       --target export --output type=local,dest=build/arm64 .
+#
+# The build runs under qemu user emulation (slow but reproducible). The base is
+# ubuntu:24.04 so the toolchain (gcc 13.3, glibc 2.39) and resulting ABI match a
+# Raspberry Pi 5 running Ubuntu 24.04 -- the binary drops straight onto the Pi
+# (it links only the system libc/libstdc++/libgomp; ggml and libparakeet are
+# static).
+#
+# By default it targets the Pi 5's Cortex-A76 (armv8.2-a + dotprod + fp16), which
+# is what ggml's quantized kernels exploit. Override --build-arg GGML_CPU_ARM_ARCH
+# for a different chip, or set it to a baseline like "armv8-a" for a portable
+# (slower) binary.
+#
+# The build context must be a checkout with the ggml submodule populated
+# (git submodule update --init --recursive third_party/ggml).
+ARG UBUNTU=24.04
+
+FROM ubuntu:${UBUNTU} AS build
+ARG GGML_CPU_ARM_ARCH=armv8.2-a+dotprod+fp16
+ARG JOBS=8
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        build-essential cmake git ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+WORKDIR /src
+COPY . .
+# The build context drops .git (see .dockerignore), so the ggml submodule's
+# .git gitlink dangles and CMake's scripts/apply_ggml_patches.sh (which uses
+# `git apply`) can't run. Re-init a throwaway git repo in the submodule tree so
+# the in-tree ggml patches (broadcast-fold sgemm, etc.) apply during configure.
+RUN rm -f third_party/ggml/.git && git -C third_party/ggml init -q
+# Static, self-contained CLI: BUILD_SHARED_LIBS=OFF keeps ggml + libparakeet in
+# the binary; GGML_NATIVE=OFF + GGML_CPU_ARM_ARCH targets the Pi's Cortex-A76
+# instead of the (x86) build host. Skip the server + tests to keep the qemu
+# build lean -- we only need parakeet-cli for the benchmark.
+RUN cmake -S . -B build/arm64 \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DBUILD_SHARED_LIBS=OFF \
+        -DGGML_NATIVE=OFF \
+        -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH} \
+        -DPARAKEET_BUILD_CLI=ON \
+        -DPARAKEET_BUILD_SERVER=OFF \
+        -DPARAKEET_BUILD_TESTS=OFF \
+    && cmake --build build/arm64 -j${JOBS} --target parakeet-cli
+
+# Thin export stage: `--output type=local` against this target writes just the
+# binary to the host (no image layers, no apt cruft).
+FROM scratch AS export
+COPY --from=build /src/build/arm64/examples/cli/parakeet-cli /