diff --git a/setup-venv-local-build.sh b/setup-venv-local-build.sh
index 84552f2..7354f27 100755
--- a/setup-venv-local-build.sh
+++ b/setup-venv-local-build.sh
@@ -96,26 +96,29 @@ ls "${VENV}/include/ctranslate2/" | head -3
 ls "${VENV}/lib/libctranslate2"* 2>/dev/null || { echo "ERROR: libctranslate2 not found in venv/lib" >&2; exit 1; }
 grep "WITH_CUDA" "${BUILD_DIR}/cmake-build/CMakeCache.txt" | grep -v "^#" || true
 
-# --- Python bindings ---
-# Always reinstall from source to ensure we use our locally built library,
-# not a PyPI wheel that may have been pulled in as a dependency.
-echo "==> removing any existing ctranslate2 install..."
+# --- faster-whisper (with all deps, including PyPI ctranslate2) ---
+# Install faster-whisper normally so all its dependencies (av, huggingface_hub, etc.)
+# are satisfied. This will pull in the PyPI ctranslate2 wheel, which we override next.
+if ! "${VENV}/bin/python3" -c "import faster_whisper" &>/dev/null 2>&1; then
+	echo "==> installing faster-whisper"
+	"${VENV}/bin/pip" install faster-whisper
+else
+	echo "==> faster-whisper already installed, skipping"
+fi
+
+# --- Python bindings (always reinstalled from source) ---
+# Override the PyPI ctranslate2 wheel pulled in above with our source-built version.
+# This is the whole point of this script: the PyPI wheel links against a fixed CUDA
+# major version (e.g. libcublas.so.12) while our build links against the system version.
+echo "==> removing PyPI ctranslate2..."
 "${VENV}/bin/pip" uninstall -y ctranslate2 2>/dev/null || true
 
-echo "==> building ctranslate2 Python bindings from source..."
+echo "==> installing source-built ctranslate2 Python bindings..."
 CT2_ROOT="${VENV}" \
 LIBRARY_PATH="${VENV}/lib:${VENV}/lib64${LIBRARY_PATH:+:${LIBRARY_PATH}}" \
 LDFLAGS="-Wl,-rpath,${VENV}/lib" \
 	"${VENV}/bin/pip" install "${BUILD_DIR}/src/python" --no-build-isolation
 
-# --- faster-whisper ---
-if ! "${VENV}/bin/python3" -c "import faster_whisper" &>/dev/null 2>&1; then
-	echo "==> installing faster-whisper"
-	"${VENV}/bin/pip" install faster-whisper --no-deps
-else
-	echo "==> faster-whisper already installed, skipping"
-fi
-
 # --- model download ---
 if [ -f "${TOKEN_FILE}" ]; then
 	export HF_TOKEN="$(cat "${TOKEN_FILE}")"