diff --git a/.github/workflows/publish-hf-wheels.yml b/.github/workflows/publish-hf-wheels.yml
index 4aaf4f4..f5f6be4 100644
--- a/.github/workflows/publish-hf-wheels.yml
+++ b/.github/workflows/publish-hf-wheels.yml
@@ -168,9 +168,71 @@ jobs:
           name: hf-wheel-${{ matrix.flavor }}-${{ matrix.arch }}
           path: wheelhouse/*.whl
 
+  build-cpu-wheels:
+    name: Linux ${{ matrix.arch }} CPU HF wheel
+    runs-on: ${{ matrix.runner }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - arch: x86_64
+            runner: ubuntu-22.04
+            manylinux_platform: manylinux_2_35_x86_64
+            build_jobs: "2"
+          - arch: aarch64
+            runner: ubuntu-22.04-arm
+            manylinux_platform: manylinux_2_35_aarch64
+            build_jobs: "2"
+    env:
+      QWENTTS_CPP_BACKEND: cpu
+      QWENTTS_CPP_BUILD_JOBS: ${{ matrix.build_jobs }}
+    steps:
+      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
+        with:
+          persist-credentials: false
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install build tools
+        run: |
+          python -m pip install --upgrade pip auditwheel build wheel
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends cmake ninja-build patchelf ca-certificates
+
+      - name: Fetch qwentts.cpp
+        run: |
+          git clone --recursive https://github.com/ServeurpersoCom/qwentts.cpp third_party/qwentts.cpp
+          git -C third_party/qwentts.cpp checkout "$QWENTTS_REF"
+          git -C third_party/qwentts.cpp submodule update --init --recursive
+
+      - name: Build wheel
+        run: |
+          python scripts/set_local_version.py cpu
+          python scripts/build_native.py \
+            --source third_party/qwentts.cpp \
+            --backend cpu \
+            --clean \
+            --cmake-arg=-G \
+            --cmake-arg=Ninja
+          python -m build --wheel
+          mkdir -p wheelhouse
+          python -m auditwheel repair \
+            --plat ${{ matrix.manylinux_platform }} \
+            -w wheelhouse \
+            dist/*.whl
+
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: hf-wheel-cpu-${{ matrix.arch }}
+          path: wheelhouse/*.whl
+
   publish-hf:
     name: Publish HF wheel index
-    needs: build-wheels
+    needs:
+      - build-wheels
+      - build-cpu-wheels
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
diff --git a/AGENTS.md b/AGENTS.md
index f757fcd..86562f1 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -20,12 +20,12 @@ The PR wheel workflow may still build CPU and CUDA 13 wheels as downloadable
 validation artifacts. Those artifacts are not uploaded to PyPI by the publish
 workflow.
 
-Additional CUDA wheels can be published to Hugging Face Hub by
+Additional backend-specific wheels can be published to Hugging Face Hub by
 `.github/workflows/publish-hf-wheels.yml`. That workflow builds local-version
-variants such as `0.2.0+cu124`, `0.2.0+cu128`, and `0.2.0+cu130`, prepares
-static `--find-links` pages, creates the public dataset repo if needed, and
-uploads the wheel index using the `HF_TOKEN` repository secret. Do not upload
-those local-version CUDA variants to PyPI.
+variants such as `0.2.0+cpu`, `0.2.0+cu124`, `0.2.0+cu128`, and
+`0.2.0+cu130`, prepares static `--find-links` pages, creates the public dataset
+repo if needed, and uploads the wheel index using the `HF_TOKEN` repository
+secret. Do not upload those local-version variants to PyPI.
 
 CUDA release wheels keep native cubins for sm_86, sm_90, and sm_120, plus PTX
 fallbacks for sm_75 and the newest supported CUDA architecture. This keeps the
diff --git a/README.md b/README.md
index 8a8b8b7..143f80b 100644
--- a/README.md
+++ b/README.md
@@ -41,11 +41,14 @@ The default PyPI package is built for CUDA 12.8:
 pip install qwentts-cpp-python
 ```
 
-Additional CUDA wheels are published to Hugging Face Hub as local-version
+Additional backend-specific wheels are published to Hugging Face Hub as local-version
 variants. Use them when the PyPI CUDA 12.8 wheel does not match the runtime or
 GPU target, for example DGX Spark / GB10 with CUDA 13:
 
 ```bash
+pip install "qwentts-cpp-python==0.2.0+cpu" \
+  -f https://huggingface.co/datasets/andimarafioti/qwentts-cpp-python-wheels/resolve/main/whl/cpu.html
+
 pip install "qwentts-cpp-python==0.2.0+cu124" \
   -f https://huggingface.co/datasets/andimarafioti/qwentts-cpp-python-wheels/resolve/main/whl/cu124.html
 
diff --git a/scripts/prepare_hf_wheel_repo.py b/scripts/prepare_hf_wheel_repo.py
index 03dd778..0ea2c6a 100644
--- a/scripts/prepare_hf_wheel_repo.py
+++ b/scripts/prepare_hf_wheel_repo.py
@@ -106,7 +106,7 @@ def main() -> int:
                 "",
                 "# qwentts-cpp-python wheels",
                 "",
-                "Optional CUDA wheel variants for `qwentts-cpp-python`.",
+                "Optional backend-specific wheel variants for `qwentts-cpp-python`.",
                 "",
                 "The default PyPI package is CUDA 12.8:",
                 "",
@@ -114,7 +114,7 @@ def main() -> int:
                 "pip install qwentts-cpp-python",
                 "```",
                 "",
-                "Install a CUDA-specific wheel from this repository with `--find-links`:",
+                "Install a backend-specific wheel from this repository with `--find-links`:",
                 "",
                 "```bash",
                 *[