diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index cadac5c9f..eca8b266a 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -11,369 +11,19 @@ on:
       - master
 
 jobs:
-  build-and-test-cpp-x86_64:
-    runs-on: ${{ matrix.os }}
-    env:
-      CT2_VERBOSE: 1
-    strategy:
-      matrix:
-        os: [ubuntu-22.04]
-        backend: [mkl, dnnl]
-
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-
-      - name: Install Intel oneAPI
-        if: startsWith(matrix.os, 'ubuntu')
-        run: |
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add *.PUB
-          sudo sh -c 'echo "deb https://apt.repos.intel.com/oneapi all main" > /etc/apt/sources.list.d/oneAPI.list'
-          sudo apt-get update
-
-      - name: Configure with MKL
-        if: startsWith(matrix.os, 'ubuntu') && matrix.backend == 'mkl'
-        env:
-          CT2_USE_MKL: 1
-          MKL_VERSION: 2023.0.0
-        run: |
-          sudo apt-get install -y intel-oneapi-mkl-devel-$MKL_VERSION
-          cmake -DCMAKE_INSTALL_PREFIX=$PWD/install -DBUILD_TESTS=ON .
-
-      - name: Configure with DNNL
-        if: startsWith(matrix.os, 'ubuntu') && matrix.backend == 'dnnl'
-        env:
-          DNNL_VERSION: 2023.0.0-25399
-        run: |
-          sudo apt-get install -y intel-oneapi-dnnl-devel=$DNNL_VERSION intel-oneapi-dnnl=$DNNL_VERSION
-          cmake -DCMAKE_INSTALL_PREFIX=$PWD/install -DBUILD_TESTS=ON -DWITH_MKL=OFF -DOPENMP_RUNTIME=COMP -DWITH_DNNL=ON .
-
-      - name: Build
-        run: |
-          make -j $(nproc) install
-
-      - name: Download test data
-        working-directory: tests/data/models
-        run: |
-          wget https://opennmt-models.s3.amazonaws.com/pi_lm_step_5000.pt
-          wget https://opennmt-models.s3.amazonaws.com/transliteration-aren-all.tar.gz
-          tar xf transliteration-aren-all.tar.gz
-
-      - name: Test MKL
-        if: matrix.backend == 'mkl'
-        env:
-          CT2_USE_MKL: 1
-        run: |
-          tests/ctranslate2_test tests/data
-      - name: Test DNNL
-        if: matrix.backend == 'dnnl'
-        run: |
-          tests/ctranslate2_test tests/data
-          
-  build-and-test-cpp-x86_64-address_sanitizer:
-    runs-on: ubuntu-22.04
-    env:
-      CT2_VERBOSE: 1
-
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-
-      - name: Install Intel oneAPI
-        run: |
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add *.PUB
-          sudo sh -c 'echo "deb https://apt.repos.intel.com/oneapi all main" > /etc/apt/sources.list.d/oneAPI.list'
-          sudo apt update
-
-      - name: Install MKL
-        env:
-          CT2_USE_MKL: 1
-          MKL_VERSION: 2023.0.0
-        run: |
-          sudo apt install -y intel-oneapi-mkl-devel-$MKL_VERSION
-
-      - name: Install Clang
-        run: |
-          sudo apt install -y clang  
-
-      - name: Configure with MKL and Clang
-        env:
-          CT2_USE_MKL: 1
-        run: |
-          cmake -DCMAKE_INSTALL_PREFIX=$PWD/install -DBUILD_TESTS=ON -DCMAKE_C_COMPILER=/usr/bin/clang -DCMAKE_CXX_COMPILER=/usr/bin/clang++ -DENABLE_ADDRESS_SANITIZER=ON -DCMAKE_BUILD_TYPE=Debug .
-
-      - name: Build
-        run: |
-          make -j $(nproc) install
-
-      - name: Download test data
-        working-directory: tests/data/models
-        run: |
-          wget https://opennmt-models.s3.amazonaws.com/pi_lm_step_5000.pt
-          wget https://opennmt-models.s3.amazonaws.com/transliteration-aren-all.tar.gz
-          tar xf transliteration-aren-all.tar.gz
-
-      - name: Test AddressSanitizer
-        env:
-          CT2_USE_MKL: 1
-        run: |
-          ASAN_OPTIONS=detect_leaks=1:print_stats=1 tests/ctranslate2_test tests/data
-          
-  build-and-test-cpp-arm64:
-    runs-on: ${{ matrix.os }}
-    env:
-      CT2_VERBOSE: 1
-    strategy:
-      matrix:
-
-        include:
-          - os: ubuntu-22.04-arm
-            backend: openblas
-          - os: macos-15
-            backend: ruy
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-
-      - name: Build with OpenBLAS and Ruy
-        if: matrix.backend == 'openblas'
-        run: |
-          wget https://github.com/xianyi/OpenBLAS/archive/v0.3.13.tar.gz
-          tar xzvf v0.3.13.tar.gz
-          cd OpenBLAS-0.3.13
-          make TARGET=ARMV8 NO_LAPACK=1 -j $(nproc)
-          sudo make PREFIX=/usr/local install -j $(nproc)
-          cd ..
-          export OpenBLAS_HOME=/usr/local
-          cmake \
-            -DOPENMP_RUNTIME=COMP \
-            -DCMAKE_INSTALL_PREFIX=$PWD/install \
-            -DWITH_MKL=OFF \
-            -DWITH_OPENBLAS=ON \
-            -DWITH_RUY=ON \
-            -DBUILD_TESTS=ON \
-            .
-          make -j $(nproc) install
-
-      - name: Build Ruy
-        if: matrix.backend == 'ruy'
-        run: |
-          CMAKE_EXTRA_OPTIONS='-DCMAKE_OSX_ARCHITECTURES=arm64 -DWITH_ACCELERATE=ON -DWITH_MKL=OFF -DOPENMP_RUNTIME=NONE -DWITH_RUY=ON'
-          cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \
-            -DCMAKE_INSTALL_RPATH_USE_LINK_PATH=ON \
-            -DCMAKE_INSTALL_PREFIX=$PWD/install \
-            -DBUILD_TESTS=ON \
-            $CMAKE_EXTRA_OPTIONS \
-            .
-          make -j $(nproc) install
-
-      - name: Download test data
-        run: |
-          wget https://opennmt-models.s3.amazonaws.com/transliteration-aren-all.tar.gz
-          tar xf transliteration-aren-all.tar.gz -C tests/data/models/
-
-      - name: Test
-        run: |
-          tests/ctranslate2_test tests/data
-
-  build-python-wheels:
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-22.04, windows-2022]
-        arch: [auto64]
-        include:
-        - os: ubuntu-22.04
-          arch: aarch64
-        - os: macos-15
-          arch: arm64
-        - os: macos-15-intel
-          arch: x86_64
-
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-
-      - uses: docker/setup-qemu-action@v2
-        if: ${{ matrix.arch == 'aarch64' }}
-        name: Set up QEMU
-
-      - name: Build wheels
-        uses: pypa/cibuildwheel@v3.2.1
-        with:
-          package-dir: python
-          output-dir: python/wheelhouse
-        env:
-          CIBW_ENVIRONMENT_PASS_LINUX: CIBW_ARCHS
-          CIBW_ENVIRONMENT_WINDOWS: CTRANSLATE2_ROOT='${{ github.workspace }}\install'
-          CIBW_ENVIRONMENT_MACOS: "CTRANSLATE2_ROOT='/usr/local' MACOSX_DEPLOYMENT_TARGET=11.00"
-          CIBW_BEFORE_ALL_LINUX: python/tools/prepare_build_environment_linux.sh
-          CIBW_BEFORE_ALL_MACOS: python/tools/prepare_build_environment_macos.sh
-          CIBW_BEFORE_ALL_WINDOWS: bash python/tools/prepare_build_environment_windows.sh
-          CIBW_BEFORE_BUILD: pip install -r python/install_requirements.txt
-          CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
-          CIBW_MANYLINUX_AARCH64_IMAGE: manylinux_2_28
-          CIBW_ARCHS: ${{ matrix.arch }}
-          CIBW_SKIP: "*-musllinux_*"
-
-      - name: Upload Python wheels
-        uses: actions/upload-artifact@v4
-        with:
-          name: python-wheels-${{ runner.os }}-${{ matrix.arch }}
-          path: python/wheelhouse
-
-  build-python-wheels-rocm:
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-24.04, windows-2025]
-
-    steps:
-      - uses: actions/checkout@v6
-        with:
-          submodules: recursive
-
-      - name: Build wheels
-        uses: pypa/cibuildwheel@v3.2.1
-        with:
-          package-dir: python
-          output-dir: python/wheelhouse
-        env:
-          CIBW_ENVIRONMENT_LINUX: ROCM_PATH=/opt/rocm LD_LIBRARY_PATH=/opt/rocm/lib/llvm/lib:$LD_LIBRARY_PATH
-          CIBW_ENVIRONMENT_WINDOWS: CTRANSLATE2_ROOT='${{ github.workspace }}\install'
-          CIBW_BEFORE_ALL_LINUX: python/tools/prepare_build_environment_linux_rocm.sh
-          CIBW_BEFORE_ALL_WINDOWS: bash python/tools/prepare_build_environment_windows_rocm.sh
-          CIBW_BEFORE_BUILD: pip install -r python/install_requirements.txt
-          CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
-          CIBW_ARCHS: auto64
-          CIBW_SKIP: "*-musllinux_*"
-          CIBW_REPAIR_WHEEL_COMMAND_LINUX: 'auditwheel repair -w {dest_dir} --exclude "/opt/rocm/lib/lib*" {wheel}'
-
-      - name: Upload Python wheels
-        uses: actions/upload-artifact@v6
-        with:
-          name: rocm-python-wheels-${{ runner.os }}
-          path: python/wheelhouse
 
-
-  # We could test the Python wheels using cibuildwheel but we prefer to run the tests outside
-  # the build environment to ensure wheels correctly embed all dependencies.
-  test-python-wheels:
-    needs: [build-python-wheels]
+  build-and-push-docker-images:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
         include:
           - os: ubuntu-22.04
-            artifact_pattern: python-wheels-Linux-auto64
-            wheel_pattern: "*cp310*manylinux*x86_64.whl"
-
+            gpu: cuda
+          - os: ubuntu-22.04
+            gpu: rocm
           - os: ubuntu-22.04-arm
-            artifact_pattern: python-wheels-Linux-aarch64
-            wheel_pattern: "*cp310*manylinux*_aarch64.whl"
-
-          - os: windows-2022
-            artifact_pattern: python-wheels-Windows-auto64
-            wheel_pattern: "*cp310*win*amd64.whl"
-
-          - os: macos-15
-            artifact_pattern: python-wheels-macOS-arm64
-            wheel_pattern: "*cp310*macosx*arm64.whl"
-
-    steps:
-      - name: Set up Python 3.10.11
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10.11"
-
-      - uses: actions/checkout@v4
-
-      - name: Prepare test environment
-        shell: bash
-        run: |
-          ./python/tools/prepare_test_environment.sh
-
-      - name: Download Python wheels
-        uses: actions/download-artifact@v4
-        with:
-          pattern: ${{ matrix.artifact_pattern }}
-          merge-multiple: true
-          path: .
-
-      - name: Install wheel
-        shell: bash
-        run: |
-          pip install ${{ matrix.wheel_pattern }}
-
-      - name: Test Python wheel
-        run: |
-          pytest -v python/tests/ --ignore=python/tests/test_opennmt_tf.py
-
-
-  check-python-style:
-    runs-on: ubuntu-latest
-
-    steps:
-      - uses: actions/checkout@v4
-
-      - name: Set up Python 3.10.11
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10.11"
-
-      - name: Install dependencies
-        run: |
-          python -m pip install black==22.* flake8==3.8.* isort==5.*
-
-      - name: Check code format with Black
-        working-directory: python
-        run: |
-          black --check .
-
-      - name: Check imports order with isort
-        working-directory: python
-        run: |
-          isort --check-only .
-
-      - name: Check code style with Flake8
-        working-directory: python
-        if: ${{ always() }}
-        run: |
-          flake8 .
-
-
-  publish-python-wheels-on-pypi:
-    if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags')
-    needs: [build-and-test-cpp-x86_64, build-and-test-cpp-arm64, build-python-wheels, test-python-wheels, check-python-style]
-    runs-on: ubuntu-22.04
-
-    steps:
-      - name: Download Python wheels
-        uses: actions/download-artifact@v4
-        with:
-          pattern: python-wheels-*
-          merge-multiple: true
-          path: .
-
-      - name: Publish Python wheels to PyPI
-        uses: pypa/gh-action-pypi-publish@release/v1
-        with:
-          user: __token__
-          password: ${{ secrets.PYPI_API_TOKEN }}
-          packages-dir: .
-
-
-  build-and-push-docker-images:
-    runs-on: ubuntu-22.04
-    strategy:
-      matrix:
-        gpu: [cuda, rocm]
+            gpu: cudaarm
+            suffix: -arm
 
     steps:
       - uses: actions/checkout@v4
@@ -405,7 +55,7 @@ jobs:
 
       - name: Build Docker images
         run: |
-          ./docker/build_all.sh latest 0 ${{ matrix.gpu }}
+          ./docker/build_all.sh latest 0 ${{ matrix.gpu }} ${{ matrix.suffix }}
 
       - name: Login to DockerHub
         if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags')
@@ -418,66 +68,4 @@ jobs:
       - name: Push Docker images
         if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags')
         run: |
-          ./docker/build_all.sh ${GITHUB_REF##*/v} 1 ${{ matrix.gpu }}
-
-
-  build-and-deploy-docs:
-    runs-on: ubuntu-latest
-    needs: [check-python-style, build-python-wheels, build-python-wheels-rocm]
-
-    permissions:
-      contents: write
-    steps:
-      - uses: actions/checkout@v4
-      - name: Set up Python 3.10.11
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10.11"
-      - name: Download CTranslate2 wheels
-        uses: actions/download-artifact@v4
-        with:
-          pattern: python-wheels-${{ runner.os }}-*
-          merge-multiple: true
-          path: .
-      - name: Install CTranslate2 wheel
-        run: |
-          pip install *cp310*manylinux*x86_64.whl
-      - name: Install dependencies to build docs
-        working-directory: docs
-        run: |
-          python -m pip install -r requirements.txt
-      - name: Build docs
-        working-directory: docs
-        run: |
-          python generate.py python
-          sphinx-build . build
-      - name: Deploy docs
-        if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags')
-        uses: JamesIves/github-pages-deploy-action@v4
-        with:
-          folder: docs/build
-          clean: true
-      - name: Download ROCm wheels and zip for release
-        if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags')
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          mkdir -p release-zips
-
-          gh run download ${{ github.run_id }} \
-            --name rocm-python-wheels-Linux \
-            --dir temp-linux
-          zip -r release-zips/rocm-python-wheels-Linux.zip temp-linux/*
-          rm -rf temp-linux
-
-          gh run download ${{ github.run_id }} \
-            --name rocm-python-wheels-Windows \
-            --dir temp-windows
-          zip -r release-zips/rocm-python-wheels-Windows.zip temp-windows/*
-          rm -rf temp-windows
-
-      - name: Create GitHub Release
-        if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags')
-        uses: softprops/action-gh-release@v2
-        with:
-          files: release-zips/*.zip
+          ./docker/build_all.sh ${GITHUB_REF##*/v} 1 ${{ matrix.gpu }} ${{ matrix.suffix }}
diff --git a/docker/Dockerfile_cuda_arm b/docker/Dockerfile_cuda_arm
new file mode 100644
index 000000000..9b242c20d
--- /dev/null
+++ b/docker/Dockerfile_cuda_arm
@@ -0,0 +1,92 @@
+FROM nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04 AS builder
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        python3-dev \
+        python3-pip \
+        wget \
+        && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /root
+
+RUN python3 -m pip --no-cache-dir install cmake==3.22.*
+
+ENV ONEDNN_VERSION=3.1.1
+RUN wget -q https://github.com/uxlfoundation/oneDNN/archive/refs/tags/v${ONEDNN_VERSION}.tar.gz && \
+    tar xf *.tar.gz && \
+    rm *.tar.gz && \
+    cd oneDNN-* && \
+    cmake -DCMAKE_BUILD_TYPE=Release -DONEDNN_LIBRARY_TYPE=STATIC -DONEDNN_BUILD_EXAMPLES=OFF -DONEDNN_BUILD_TESTS=OFF -DONEDNN_ENABLE_WORKLOAD=INFERENCE -DONEDNN_ENABLE_PRIMITIVE="CONVOLUTION;REORDER" -DONEDNN_BUILD_GRAPH=OFF . && \
+    make -j$(nproc) install && \
+    cd .. && \
+    rm -r oneDNN-*
+
+ENV OPENMPI_VERSION=4.1.6
+RUN wget -q https://download.open-mpi.org/release/open-mpi/v4.1/openmpi-${OPENMPI_VERSION}.tar.bz2 && \
+    tar xf *.tar.bz2 && \
+    rm *.tar.bz2 && \
+    cd openmpi-* && \
+    ./configure && \
+    make -j$(nproc) install && \
+    cd .. && \
+    rm -r openmpi-*
+
+COPY third_party third_party
+COPY cli cli
+COPY include include
+COPY src src
+COPY cmake cmake
+COPY python python
+COPY CMakeLists.txt .
+
+ARG CXX_FLAGS
+ARG CUDA_NVCC_FLAGS
+ENV CUDA_NVCC_FLAGS=${CUDA_NVCC_FLAGS:-"-Xfatbin=-compress-all"}
+ARG CUDA_ARCH_LIST
+ENV CUDA_ARCH_LIST=${CUDA_ARCH_LIST:-"Common"}
+ENV CTRANSLATE2_ROOT=/opt/ctranslate2
+ENV LD_LIBRARY_PATH=/usr/local/lib/:${LD_LIBRARY_PATH}
+
+RUN mkdir build_tmp && \
+    cd build_tmp && \
+    cmake -DCMAKE_INSTALL_PREFIX=${CTRANSLATE2_ROOT} \
+          -DWITH_CUDA=ON -DWITH_CUDNN=ON -DWITH_MKL=OFF -DWITH_DNNL=ON -DOPENMP_RUNTIME=COMP \
+          -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="${CXX_FLAGS}" \
+          -DCUDA_NVCC_FLAGS="${CUDA_NVCC_FLAGS}" -DCUDA_ARCH_LIST="${CUDA_ARCH_LIST}" -DWITH_TENSOR_PARALLEL=ON .. && \
+    VERBOSE=1 make -j$(nproc) install
+
+ENV LANG=en_US.UTF-8
+COPY README.md .
+
+RUN cd python && \
+    python3 -m pip --no-cache-dir install -r install_requirements.txt && \
+    python3 setup.py bdist_wheel --dist-dir $CTRANSLATE2_ROOT
+
+FROM nvidia/cuda:12.8.1-base-ubuntu22.04
+
+# We remove the cuda-compat package because it conflicts with the CUDA Enhanced Compatibility.
+# See e.g. https://github.com/NVIDIA/nvidia-docker/issues/1515
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        libcublas-12-8 \
+        libcudnn9-cuda-12 \
+        libnccl2 \
+        libopenmpi3 \
+        openmpi-bin \
+        libgomp1 \
+        python3-pip \
+        && \
+    apt-get purge -y cuda-compat-12-8 || true && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+ENV CTRANSLATE2_ROOT=/opt/ctranslate2
+ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CTRANSLATE2_ROOT/lib
+
+COPY --from=builder $CTRANSLATE2_ROOT $CTRANSLATE2_ROOT
+RUN python3 -m pip --no-cache-dir install $CTRANSLATE2_ROOT/*.whl && \
+    rm $CTRANSLATE2_ROOT/*.whl
+
+ENTRYPOINT ["/opt/ctranslate2/bin/ct2-translator"]
diff --git a/docker/build_all.sh b/docker/build_all.sh
index f05ea99e8..03cf31e67 100755
--- a/docker/build_all.sh
+++ b/docker/build_all.sh
@@ -21,7 +21,8 @@ cd $ROOT_DIR
 VERSION=${1:-latest}
 PUSH=${2:-0}
 GPU=${3:-cuda}
-IMAGE=ghcr.io/opennmt/ctranslate2
+SUFFIX=${4}
+IMAGE=ghcr.io/nicolasbfr/ctranslate2
 
 build()
 {
@@ -29,8 +30,8 @@ build()
     IMAGE_NAME=$2
     BUILD_ARGS=${3:-}
 
-    LATEST=$IMAGE:latest-$IMAGE_NAME
-    TAGGED=$IMAGE:$VERSION-$IMAGE_NAME
+    LATEST=$IMAGE:latest-$IMAGE_NAME$SUFFIX
+    TAGGED=$IMAGE:$VERSION-$IMAGE_NAME$SUFFIX
     docker build --pull $BUILD_ARGS -t $LATEST -f docker/$DOCKERFILE .
     if [ $PUSH -eq 1 ]; then
         docker push $LATEST
@@ -45,6 +46,8 @@ build()
 
 if [ "$GPU" == "rocm" ]; then
     build Dockerfile_rocm ubuntu22.04-rocm7.2
+elif [ "$GPU" == "cudaarm" ]; then
+    build Dockerfile_cuda_arm ubuntu22.04-cuda12.8
 else
     build Dockerfile ubuntu22.04-cuda12.8
 fi