diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 299639c..4af0f0c 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -8,7 +8,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  ACESTEP_CPP_REPO: https://github.com/ServeurpersoCom/acestep.cpp.git
+  ACESTEP_CPP_REPO: https://github.com/audiohacking/acestep.cpp.git
 
 jobs:
   # ─────────────────────────────────────────────
@@ -76,10 +76,14 @@ jobs:
           retention-days: 7
 
   # ─────────────────────────────────────────────
-  # 3. Build acestep.cpp with hardware detection
+  # 3. Test build.sh (the deployment build script)
+  #    Validates that build.sh correctly clones
+  #    acestep.cpp, compiles it, and installs the
+  #    binaries to bin/ — exactly as it will run
+  #    on the end user's machine at first launch.
   # ─────────────────────────────────────────────
-  build-acestep-cpp:
-    name: Build acestep.cpp (${{ matrix.label }})
+  test-build-script:
+    name: Test build.sh (${{ matrix.label }})
     runs-on: ${{ matrix.runner }}
     permissions:
       contents: read
@@ -97,117 +101,52 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      # ── System dependencies ─────────────────────────────────────────────
+      # ── Install build prerequisites (same as a user would need) ──────────
       - name: Install build tools (Linux)
         if: matrix.os == 'linux'
         run: |
           sudo apt-get update -qq
-          sudo apt-get install -y cmake ninja-build libssl-dev pkg-config libopenblas-dev
+          sudo apt-get install -y cmake libssl-dev pkg-config
 
       - name: Install build tools (macOS)
         if: matrix.os == 'macos'
-        run: brew install cmake ninja
+        run: brew install cmake
 
-      # ── Clone acestep.cpp with submodules ────────────────────────────────
-      - name: Clone acestep.cpp
-        run: |
-          git clone --depth 1 ${{ env.ACESTEP_CPP_REPO }} acestep.cpp
-          cd acestep.cpp
-          git submodule update --init --recursive --depth 1
-
-      # ── Detect hardware and set cmake flags ──────────────────────────────
-      # Follows the detection approach documented in the acestep.cpp README:
-      #   CUDA (NVIDIA):  -DGGML_CUDA=ON
-      #   ROCm/HIP (AMD): -DGGML_HIP=ON
-      #   Vulkan:         -DGGML_VULKAN=ON
-      #   OpenBLAS (CPU): -DGGML_BLAS=ON
-      #   macOS:          Metal + Accelerate BLAS auto-detected by cmake
-      - name: Detect hardware and set cmake flags
-        id: detect
-        shell: bash
-        run: |
-          FLAGS=""
-          if [[ "$(uname)" == "Darwin" ]]; then
-            echo "Platform: macOS — Metal and Accelerate BLAS auto-detected by cmake"
-          else
-            # CUDA (NVIDIA)
-            if command -v nvcc &>/dev/null || [ -d /usr/local/cuda ] || [ -d /usr/cuda ]; then
-              FLAGS="$FLAGS -DGGML_CUDA=ON"
-              echo "Detected: CUDA"
-            fi
-            # ROCm/HIP (AMD)
-            if command -v hipcc &>/dev/null || [ -d /opt/rocm ]; then
-              FLAGS="$FLAGS -DGGML_HIP=ON"
-              echo "Detected: ROCm/HIP"
-            fi
-            # Vulkan
-            if pkg-config --exists vulkan 2>/dev/null; then
-              FLAGS="$FLAGS -DGGML_VULKAN=ON"
-              echo "Detected: Vulkan"
-            fi
-            # OpenBLAS (recommended for CPU-only machines)
-            if pkg-config --exists openblas 2>/dev/null; then
-              FLAGS="$FLAGS -DGGML_BLAS=ON"
-              echo "Detected: OpenBLAS"
-            fi
-            [ -z "$FLAGS" ] && echo "No accelerator detected — CPU-only build"
-          fi
-          echo "flags=$FLAGS" >> "$GITHUB_OUTPUT"
-          echo "cmake flags: $FLAGS"
-
-      # ── Build ────────────────────────────────────────────────────────────
-      - name: Cache acestep.cpp build artifacts
+      # ── Cache the acestep.cpp build directory ────────────────────────────
+      - name: Cache acestep.cpp build
         uses: actions/cache@v4
         with:
-          path: acestep.cpp/build
-          key: acestep-ci-${{ matrix.label }}-${{ hashFiles('acestep.cpp/CMakeLists.txt') }}
-          restore-keys: acestep-ci-${{ matrix.label }}-
-
-      - name: Build acestep.cpp
-        run: |
-          cmake -S acestep.cpp -B acestep.cpp/build \
-            -G Ninja \
-            -DCMAKE_BUILD_TYPE=Release \
-            ${{ steps.detect.outputs.flags }}
-          cmake --build acestep.cpp/build --parallel
-
-      # ── Verify and collect binaries ──────────────────────────────────────
-      - name: Verify binaries
+          path: |
+            acestep.cpp
+            bin
+          key: build-script-${{ matrix.label }}-${{ hashFiles('build.sh') }}-${{ env.ACESTEP_CPP_REPO }}
+          restore-keys: build-script-${{ matrix.label }}-
+
+      # ── Run the real deployment build script (CPU-only on CI runners) ────
+      - name: Run build.sh
+        run: bash build.sh --cpu
+
+      # ── Verify expected binaries were installed to bin/ ──────────────────
+      - name: Verify installed binaries
         shell: bash
         run: |
           ALL_OK=1
           for bin in ace-qwen3 dit-vae neural-codec; do
-            BIN_PATH=$(find acestep.cpp/build -name "$bin" -type f 2>/dev/null | head -1)
-            if [ -f "$BIN_PATH" ]; then
-              echo "✅ $bin → $BIN_PATH"
+            if [ -x "bin/$bin" ]; then
+              echo "✅ bin/$bin"
             else
-              echo "⚠️  $bin not found in build tree"
+              echo "⚠️  bin/$bin not found or not executable"
               ALL_OK=0
             fi
           done
-          if [ "$ALL_OK" = "1" ]; then
-            echo "All binaries built successfully"
-          else
-            echo "One or more binaries missing — check the build output above"
-            exit 1
-          fi
-
-      - name: Collect binaries
-        shell: bash
-        run: |
-          mkdir -p ci-bin
-          for bin in ace-qwen3 dit-vae neural-codec; do
-            BIN_PATH=$(find acestep.cpp/build -name "$bin" -type f 2>/dev/null | head -1)
-            if [ -f "$BIN_PATH" ]; then
-              cp "$BIN_PATH" "ci-bin/$bin"
-            fi
-          done
-          ls -lh ci-bin/
+          [ "$ALL_OK" = "1" ] || { echo "One or more binaries missing"; exit 1; }
 
-      - uses: actions/upload-artifact@v4
+      - name: Upload binaries artifact
+        uses: actions/upload-artifact@v4
         with:
-          name: acestep-cpp-${{ matrix.label }}
-          path: ci-bin/
+          name: ci-bin-${{ matrix.label }}
+          path: bin/
           if-no-files-found: warn
           retention-days: 7
 
+
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 73d2ff2..9a54746 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -14,7 +14,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  ACESTEP_CPP_REPO: https://github.com/ServeurpersoCom/acestep.cpp.git
+  ACESTEP_CPP_REPO: https://github.com/audiohacking/acestep.cpp.git
   NODE_VERSION: "20"
 
 jobs:
@@ -75,133 +75,32 @@ jobs:
           retention-days: 1
 
   # ────────────────────────────────────────────────────────────────────────────
-  # 3. Per-platform: build acestep.cpp, assemble bundle
+  # 3. Per-platform: assemble the bundle with build + model scripts
+  #    acestep.cpp is NOT compiled here — it is compiled on the user's machine
+  #    by build.sh / build.bat on first launch, which detects the actual GPU.
   # ────────────────────────────────────────────────────────────────────────────
   build-platform:
-    name: Build ${{ matrix.label }}
+    name: Bundle ${{ matrix.label }}
     needs: [build-frontend, build-server]
-    runs-on: ${{ matrix.runner }}
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
     strategy:
       fail-fast: false
       matrix:
         include:
-          # Linux x86_64 — detect available accelerators at build time
           - label: linux-x64
-            runner: ubuntu-22.04
             os: linux
-
-          # Linux ARM64 — detect available accelerators at build time
           - label: linux-arm64
-            runner: ubuntu-22.04-arm
             os: linux
-
-          # macOS Apple Silicon — Metal + Accelerate BLAS auto-detected by cmake
           - label: macos-arm64
-            runner: macos-14
             os: macos
+          - label: windows-x64
+            os: windows
 
     steps:
       - uses: actions/checkout@v4
 
-      # ── System dependencies ─────────────────────────────────────────────
-      - name: Install build tools (Linux)
-        if: matrix.os == 'linux'
-        run: |
-          sudo apt-get update -qq
-          sudo apt-get install -y cmake ninja-build libssl-dev pkg-config libopenblas-dev
-
-      - name: Install build tools (macOS)
-        if: matrix.os == 'macos'
-        run: brew install cmake ninja
-
-      # ── Clone & build acestep.cpp ───────────────────────────────────────
-      - name: Clone acestep.cpp
-        run: |
-          git clone --depth 1 ${{ env.ACESTEP_CPP_REPO }} acestep.cpp
-          cd acestep.cpp
-          git submodule update --init --recursive --depth 1
-
-      # ── Detect hardware and set cmake flags ──────────────────────────────
-      # Follows the detection approach documented in the acestep.cpp README:
-      #   CUDA (NVIDIA):  -DGGML_CUDA=ON
-      #   ROCm/HIP (AMD): -DGGML_HIP=ON
-      #   Vulkan:         -DGGML_VULKAN=ON
-      #   OpenBLAS (CPU): -DGGML_BLAS=ON
-      #   macOS:          Metal + Accelerate BLAS auto-detected by cmake
-      - name: Detect hardware and set cmake flags
-        id: detect
-        shell: bash
-        run: |
-          FLAGS=""
-          if [[ "$(uname)" == "Darwin" ]]; then
-            echo "Platform: macOS — Metal and Accelerate BLAS auto-detected by cmake"
-          else
-            # CUDA (NVIDIA)
-            if command -v nvcc &>/dev/null || [ -d /usr/local/cuda ] || [ -d /usr/cuda ]; then
-              FLAGS="$FLAGS -DGGML_CUDA=ON"
-              echo "Detected: CUDA"
-            fi
-            # ROCm/HIP (AMD)
-            if command -v hipcc &>/dev/null || [ -d /opt/rocm ]; then
-              FLAGS="$FLAGS -DGGML_HIP=ON"
-              echo "Detected: ROCm/HIP"
-            fi
-            # Vulkan
-            if pkg-config --exists vulkan 2>/dev/null; then
-              FLAGS="$FLAGS -DGGML_VULKAN=ON"
-              echo "Detected: Vulkan"
-            fi
-            # OpenBLAS (recommended for CPU-only machines)
-            if pkg-config --exists openblas 2>/dev/null; then
-              FLAGS="$FLAGS -DGGML_BLAS=ON"
-              echo "Detected: OpenBLAS"
-            fi
-            [ -z "$FLAGS" ] && echo "No accelerator detected — CPU-only build"
-          fi
-          echo "flags=$FLAGS" >> "$GITHUB_OUTPUT"
-          echo "cmake flags: $FLAGS"
-
-      - name: Cache acestep.cpp build artifacts
-        uses: actions/cache@v4
-        with:
-          path: acestep.cpp/build
-          key: acestep-${{ matrix.label }}-${{ hashFiles('acestep.cpp/CMakeLists.txt') }}
-          restore-keys: acestep-${{ matrix.label }}-
-
-      - name: Build acestep.cpp
-        run: |
-          cmake -S acestep.cpp -B acestep.cpp/build \
-            -G Ninja \
-            -DCMAKE_BUILD_TYPE=Release \
-            ${{ steps.detect.outputs.flags }}
-          cmake --build acestep.cpp/build --parallel
-
-      # Locate the three generated binaries
-      - name: Locate binaries
-        id: locate-bins
-        shell: bash
-        run: |
-          locate_bin() {
-            local name="$1"
-            local found
-            found=$(find acestep.cpp/build -name "$name" -type f 2>/dev/null | head -1)
-            if [ -n "$found" ]; then
-              echo "Found $name → $found" >&2
-              printf '%s' "$found"
-            else
-              echo "⚠️  $name not found in build tree" >&2
-              printf ''
-            fi
-          }
-
-          LM_BIN=$(locate_bin "ace-qwen3")
-          DIT_VAE_BIN=$(locate_bin "dit-vae")
-          NEURAL_CODEC_BIN=$(locate_bin "neural-codec")
-
-          echo "lm_bin=$LM_BIN"               >> "$GITHUB_OUTPUT"
-          echo "dit_vae_bin=$DIT_VAE_BIN"     >> "$GITHUB_OUTPUT"
-          echo "neural_codec_bin=$NEURAL_CODEC_BIN" >> "$GITHUB_OUTPUT"
-
       # ── Download pre-built UI artifacts ─────────────────────────────────
       - uses: actions/download-artifact@v4
         with:
@@ -220,7 +119,7 @@ jobs:
           set -e
           APP=acestep-cpp-ui
           mkdir -p "$APP/bin" "$APP/dist" "$APP/server/dist" "$APP/server/node_modules" \
-                   "$APP/data" "$APP/public/audio" "$APP/logs"
+                   "$APP/data" "$APP/public/audio" "$APP/logs" "$APP/models"
 
           # Node.js server (compiled JS + production node_modules)
           cp -r _release/server-bundle/dist/.          "$APP/server/dist/"
@@ -230,162 +129,218 @@ jobs:
           # React frontend (served by the API server)
           cp -r _release/dist/. "$APP/dist/"
 
-          # acestep.cpp binaries — all three are needed at runtime:
-          #   ace-qwen3:    LLM (generates lyrics + audio codes from caption)
-          #   dit-vae:      DiT + VAE (synthesises audio from audio codes)
-          #   neural-codec: VAE encode/decode (auxiliary)
-          copy_bin() {
-            local name="$1"
-            local src="$2"
-            if [ -f "$src" ]; then
-              cp "$src" "$APP/bin/$name"
-              chmod +x "$APP/bin/$name"
-              echo "✅ Bundled $name"
-            else
-              echo "⚠️  $name not found — bundle will need the binary added manually"
-            fi
-          }
-          copy_bin "ace-qwen3"    "${{ steps.locate-bins.outputs.lm_bin }}"
-          copy_bin "dit-vae"      "${{ steps.locate-bins.outputs.dit_vae_bin }}"
-          copy_bin "neural-codec" "${{ steps.locate-bins.outputs.neural_codec_bin }}"
-
           # .env template
           cp .env.example "$APP/.env.example"
 
-          # start.sh
+          # Build scripts — compile acestep.cpp with GPU detection on first run
+          cp build.sh   "$APP/build.sh"
+          chmod +x      "$APP/build.sh"
+          cp build.bat  "$APP/build.bat"
+
+          # Model download scripts
+          cp models.sh  "$APP/models.sh"
+          chmod +x      "$APP/models.sh"
+          cp models.bat "$APP/models.bat"
+
+          # ── start.sh (Linux / macOS) ──────────────────────────────────────
           cat > "$APP/start.sh" << 'SH'
-          #!/bin/bash
-          # ACE-Step UI — quick start
-          set -e
-          DIR="$(cd "$(dirname "$0")" && pwd)"
-          cd "$DIR"
-
-          # Load .env if present
-          [ -f .env ] && export $(grep -v '^#' .env | grep -v '^\s*$' | xargs)
-
-          # Binaries bundled under bin/
-          : "${ACE_QWEN3_BIN:=$DIR/bin/ace-qwen3}"
-          : "${DIT_VAE_BIN:=$DIR/bin/dit-vae}"
-
-          # Models — download with ./models.sh if not present
-          : "${MODELS_DIR:=$DIR/models}"
-          : "${PORT:=3001}"
-
-          mkdir -p logs data public/audio models
-
-          echo "Starting API server on port $PORT ..."
-          ACE_QWEN3_BIN="$ACE_QWEN3_BIN" \
-          DIT_VAE_BIN="$DIT_VAE_BIN" \
-          MODELS_DIR="$MODELS_DIR" \
-          PORT="$PORT" \
-          DATABASE_PATH="$DIR/data/acestep.db" \
-          AUDIO_DIR="$DIR/public/audio" \
-          FRONTEND_URL="http://localhost:$PORT" \
-            node "$DIR/server/dist/index.js" &
-          NODE_PID=$!
-
-          sleep 2
-          echo ""
-          echo "ACE-Step UI → http://localhost:$PORT"
-          echo ""
-          trap 'kill $NODE_PID 2>/dev/null; exit 0' INT TERM
-          wait $NODE_PID
-          SH
+#!/bin/bash
+# ACE-Step UI — start
+# On first launch: compiles acestep.cpp for your GPU and downloads base models.
+set -e
+DIR="$(cd "$(dirname "$0")" && pwd)"
+cd "$DIR"
+
+# Load .env if present
+[ -f .env ] && export $(grep -v '^#' .env | grep -v '^\s*$' | xargs)
+
+# ── First-run: compile acestep.cpp ───────────────────────────────────────────
+if [ ! -x "$DIR/bin/ace-qwen3" ] || [ ! -x "$DIR/bin/dit-vae" ]; then
+  echo "═══════════════════════════════════════════════════════"
+  echo "  First run: building acestep.cpp for your hardware"
+  echo "  GPU support (CUDA / ROCm / Vulkan / Metal) is"
+  echo "  detected automatically from your system."
+  echo "  This takes a few minutes and only happens once."
+  echo "═══════════════════════════════════════════════════════"
+  echo ""
+  bash "$DIR/build.sh"
+  echo ""
+fi
+
+# ── First-run: download models ───────────────────────────────────────────────
+: "${MODELS_DIR:=$DIR/models}"
+if ! ls "$MODELS_DIR"/*.gguf &>/dev/null 2>&1; then
+  echo "═══════════════════════════════════════════════════════"
+  echo "  First run: downloading base GGUF models (~8 GB)"
+  echo "  VAE + Text Encoder + LM-4B + DiT-Turbo (Q8_0)"
+  echo "  Requires: curl or wget"
+  echo "═══════════════════════════════════════════════════════"
+  echo ""
+  bash "$DIR/models.sh" --dir "$MODELS_DIR"
+  echo ""
+fi
+
+# ── Start the server ─────────────────────────────────────────────────────────
+: "${ACE_QWEN3_BIN:=$DIR/bin/ace-qwen3}"
+: "${DIT_VAE_BIN:=$DIR/bin/dit-vae}"
+: "${PORT:=3001}"
+
+mkdir -p logs data public/audio
+
+echo "Starting ACE-Step UI on port $PORT ..."
+ACE_QWEN3_BIN="$ACE_QWEN3_BIN" \
+DIT_VAE_BIN="$DIT_VAE_BIN" \
+MODELS_DIR="$MODELS_DIR" \
+PORT="$PORT" \
+DATABASE_PATH="$DIR/data/acestep.db" \
+AUDIO_DIR="$DIR/public/audio" \
+FRONTEND_URL="http://localhost:$PORT" \
+  node "$DIR/server/dist/index.js" &
+NODE_PID=$!
+
+sleep 2
+echo ""
+echo "ACE-Step UI → http://localhost:$PORT"
+echo ""
+trap 'kill $NODE_PID 2>/dev/null; exit 0' INT TERM
+wait $NODE_PID
+SH
           chmod +x "$APP/start.sh"
 
-          # start.bat (Windows — needs WSL or separate build for the binary)
+          # ── start.bat (Windows) ───────────────────────────────────────────
           cat > "$APP/start.bat" << 'BAT'
-          @echo off
-          setlocal
-          set DIR=%~dp0
-          if "%ACE_QWEN3_BIN%"=="" set ACE_QWEN3_BIN=%DIR%bin\ace-qwen3.exe
-          if "%DIT_VAE_BIN%"==""   set DIT_VAE_BIN=%DIR%bin\dit-vae.exe
-          if "%MODELS_DIR%"==""    set MODELS_DIR=%DIR%models
-          if "%PORT%"==""          set PORT=3001
-          mkdir logs 2>nul & mkdir data 2>nul & mkdir public\audio 2>nul & mkdir models 2>nul
-          set DATABASE_PATH=%DIR%data\acestep.db
-          set AUDIO_DIR=%DIR%public\audio
-          set FRONTEND_URL=http://localhost:%PORT%
-          node "%DIR%server\dist\index.js"
-          BAT
-
-          # Release README
+@echo off
+:: ACE-Step UI — start
+:: On first launch: compiles acestep.cpp for your GPU and downloads base models.
+setlocal EnableDelayedExpansion
+set DIR=%~dp0
+cd /d "%DIR%"
+
+:: Load .env if present
+if exist ".env" (
+    for /f "usebackq tokens=* delims=" %%l in (".env") do (
+        set line=%%l
+        if not "!line:~0,1!"=="#" if not "!line!"=="" set %%l
+    )
+)
+
+:: First-run: compile acestep.cpp
+if not exist "%DIR%bin\ace-qwen3.exe" (
+    echo ═══════════════════════════════════════════════════════
+    echo   First run: building acestep.cpp for your hardware
+    echo   GPU support ^(CUDA / Vulkan^) is detected automatically.
+    echo   This takes a few minutes and only happens once.
+    echo ═══════════════════════════════════════════════════════
+    echo.
+    call "%DIR%build.bat"
+    if %ERRORLEVEL% NEQ 0 ( echo Build failed. See output above. & pause & exit /b 1 )
+    echo.
+)
+
+:: First-run: download models
+if "%MODELS_DIR%"=="" set MODELS_DIR=%DIR%models
+dir /b "%MODELS_DIR%\*.gguf" >nul 2>&1
+if %ERRORLEVEL% NEQ 0 (
+    echo ═══════════════════════════════════════════════════════
+    echo   First run: downloading base GGUF models ~8 GB
+    echo   VAE + Text Encoder + LM-4B + DiT-Turbo ^(Q8_0^)
+    echo ═══════════════════════════════════════════════════════
+    echo.
+    call "%DIR%models.bat" --dir "%MODELS_DIR%"
+    echo.
+)
+
+:: Start the server
+if "%ACE_QWEN3_BIN%"=="" set ACE_QWEN3_BIN=%DIR%bin\ace-qwen3.exe
+if "%DIT_VAE_BIN%"==""   set DIT_VAE_BIN=%DIR%bin\dit-vae.exe
+if "%PORT%"==""           set PORT=3001
+mkdir logs 2>nul & mkdir data 2>nul & mkdir public\audio 2>nul
+
+echo Starting ACE-Step UI on port %PORT% ...
+set DATABASE_PATH=%DIR%data\acestep.db
+set AUDIO_DIR=%DIR%public\audio
+set FRONTEND_URL=http://localhost:%PORT%
+node "%DIR%server\dist\index.js"
+BAT
+
+          # ── Release README ────────────────────────────────────────────────
           cat > "$APP/README.md" << 'MD'
-          # ACE-Step UI
-
-          Self-contained release bundle — no Python required.  
-          Music generation is powered by [acestep.cpp](https://github.com/ServeurpersoCom/acestep.cpp).
-
-          ## Requirements
-
-          - **Node.js ≥ 20** (runtime for the API server)
-          - **GGUF model files** — download with `./models.sh` (requires `curl`)
-
-          ## Quick Start
-
-          ```bash
-          # 1. Download models (~7.7 GB for Q8_0 defaults)
-          ./models.sh
-
-          # 2. Start the server
-          ./start.sh
-          # → open http://localhost:3001
-          ```
-
-          Or copy `.env.example` → `.env`, set `MODELS_DIR` to your model directory, then run `./start.sh`.
-
-          ## Binaries
-
-          | Binary | Role |
-          |--------|------|
-          | `bin/ace-qwen3` | LLM — generates lyrics and audio codes from caption |
-          | `bin/dit-vae` | DiT + VAE — synthesises stereo 48 kHz WAV from audio codes |
-          | `bin/neural-codec` | VAE encode/decode (auxiliary) |
-
-          ## Environment Variables
-
-          | Variable | Default | Description |
-          |----------|---------|-------------|
-          | `ACE_QWEN3_BIN` | `./bin/ace-qwen3` | Path to the ace-qwen3 LLM binary |
-          | `DIT_VAE_BIN` | `./bin/dit-vae` | Path to the dit-vae binary |
-          | `MODELS_DIR` | `./models` | Directory containing GGUF model files |
-          | `PORT` | `3001` | API + UI server port |
-          | `AUDIO_DIR` | `./public/audio` | Where generated audio is stored |
-          | `JWT_SECRET` | built-in | Change for multi-user deployments |
-
-          ## GPU Acceleration
-
-          This bundle is built with auto-detected hardware support.
-          For an explicit GPU build, rebuild `acestep.cpp` from source:
-
-          ```bash
-          # CUDA (Linux/Windows)
-          cmake -S acestep.cpp -B build -DGGML_CUDA=ON && cmake --build build -j$(nproc)
-
-          # Metal (macOS — auto-detected, nothing extra needed)
-          cmake -S acestep.cpp -B build && cmake --build build -j$(nproc)
-
-          # ROCm/HIP (AMD)
-          cmake -S acestep.cpp -B build -DGGML_HIP=ON && cmake --build build -j$(nproc)
-          ```
-
-          Then replace the binaries in `bin/` with the newly built ones.
-          MD
-
-          echo "Bundle ready:"
-          ls -lh "$APP/bin/" 2>/dev/null || echo "  (no binaries)"
-
-      # ── Strip debug symbols ──────────────────────────────────────────────
-      - name: Strip binary (Linux)
-        if: matrix.os == 'linux'
-        run: strip --strip-unneeded acestep-cpp-ui/bin/* 2>/dev/null || true
-
-      - name: Strip binary (macOS)
-        if: matrix.os == 'macos'
-        run: strip acestep-cpp-ui/bin/* 2>/dev/null || true
+# ACE-Step UI
+
+Self-contained release bundle — no Python required.  
+Music generation is powered by [acestep.cpp](https://github.com/audiohacking/acestep.cpp).
+
+## Requirements
+
+- **Node.js ≥ 20** — runtime for the API server ([nodejs.org](https://nodejs.org))
+- **cmake + git** — required by `build.sh` / `build.bat` to compile `acestep.cpp`
+- **curl or wget** — required by `models.sh` to download model files
+- **~8 GB free disk space** — for the default `Q8_0` model set
+
+## Quick Start
+
+```bash
+# Just run start.sh — it handles everything on first launch:
+#   1. Detects your GPU (CUDA / ROCm / Vulkan / Metal) and compiles acestep.cpp
+#   2. Downloads the base GGUF models (~8 GB)
+#   3. Starts the UI on http://localhost:3001
+./start.sh
+```
+
+On Windows:
+```bat
+start.bat
+```
+
+## GPU Acceleration
+
+`build.sh` / `build.bat` automatically detect your GPU on first launch:
+
+| GPU | Flag used |
+|-----|-----------|
+| NVIDIA (CUDA) | `-DGGML_CUDA=ON` |
+| AMD (ROCm/HIP) | `-DGGML_HIP=ON` |
+| Any (Vulkan) | `-DGGML_VULKAN=ON` |
+| macOS (Metal) | auto-detected by cmake |
+| CPU only | no flags |
+
+To force a specific backend, run the build script manually:
+```bash
+./build.sh --cuda    # Force CUDA
+./build.sh --rocm    # Force ROCm/HIP
+./build.sh --vulkan  # Force Vulkan
+./build.sh --cpu     # CPU-only
+```
+
+## Environment Variables
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `ACE_QWEN3_BIN` | `./bin/ace-qwen3` | Path to the ace-qwen3 LLM binary |
+| `DIT_VAE_BIN` | `./bin/dit-vae` | Path to the dit-vae binary |
+| `MODELS_DIR` | `./models` | Directory containing GGUF model files |
+| `PORT` | `3001` | API + UI server port |
+| `AUDIO_DIR` | `./public/audio` | Where generated audio is stored |
+| `JWT_SECRET` | built-in | Change for multi-user deployments |
+
+Copy `.env.example` → `.env` to customise these.
+
+## Model Management
+
+Download models manually (useful for selecting a different quantisation):
+```bash
+./models.sh              # Default: Q8_0 essentials (~8 GB)
+./models.sh --quant Q5_K_M  # Smaller, slightly lower quality
+./models.sh --all        # Every model and quantisation
+```
+MD
+
+          echo "Bundle assembled for ${{ matrix.label }}:"
+          ls -lh "$APP/"
 
       # ── Package ──────────────────────────────────────────────────────────
-      - name: Create tarball
+      - name: Create archive (tar.gz)
+        if: matrix.os != 'windows'
         shell: bash
         run: |
           TAG="${GITHUB_REF_NAME:-${{ github.event.inputs.tag }}}"
@@ -395,6 +350,17 @@ jobs:
           echo "ARCHIVE=$ARCHIVE" >> "$GITHUB_ENV"
           ls -lh "$ARCHIVE"
 
+      - name: Create archive (zip — Windows)
+        if: matrix.os == 'windows'
+        shell: bash
+        run: |
+          TAG="${GITHUB_REF_NAME:-${{ github.event.inputs.tag }}}"
+          [[ "$TAG" != v* ]] && TAG="${{ github.event.inputs.tag || 'v0.0.0-dev' }}"
+          ARCHIVE="acestep-cpp-ui-${TAG}-${{ matrix.label }}.zip"
+          zip -qr "$ARCHIVE" acestep-cpp-ui/
+          echo "ARCHIVE=$ARCHIVE" >> "$GITHUB_ENV"
+          ls -lh "$ARCHIVE"
+
       - uses: actions/upload-artifact@v4
         with:
           name: release-${{ matrix.label }}
@@ -454,25 +420,33 @@ jobs:
             ## ACE-Step UI ${{ steps.tag.outputs.tag }}
 
             Self-contained bundle — **no Python required**.  
-            The Node.js API server spawns `acestep-generate` directly; no extra processes to manage.
+            Music generation is powered by [acestep.cpp](https://github.com/audiohacking/acestep.cpp).
+
+            > **GPU acceleration is built on your machine** — `start.sh` / `start.bat` automatically
+            > detects your GPU (CUDA / ROCm / Vulkan / Metal) and compiles `acestep.cpp` on first launch.
+            > Models are also downloaded automatically on first run.
 
             ### Platforms
             | File | OS | Arch |
             |------|----|------|
-            | `*-linux-x64.tar.gz`  | Linux  | x86_64       |
-            | `*-linux-arm64.tar.gz`| Linux  | ARM64        |
-            | `*-macos-arm64.tar.gz`| macOS  | Apple Silicon|
+            | `*-linux-x64.tar.gz`   | Linux   | x86\_64        |
+            | `*-linux-arm64.tar.gz` | Linux   | ARM64          |
+            | `*-macos-arm64.tar.gz` | macOS   | Apple Silicon  |
+            | `*-windows-x64.zip`    | Windows | x86\_64        |
+
+            ### Requirements
+            - **Node.js ≥ 20** — [nodejs.org](https://nodejs.org)
+            - **cmake + git** — to compile `acestep.cpp` on first launch
+            - **curl or wget** — to download GGUF model files on first launch
 
             ### Quick Start
             ```bash
             tar xzf acestep-cpp-ui-*-linux-x64.tar.gz
             cd acestep-cpp-ui
-            export ACESTEP_MODEL=/path/to/model.gguf
-            ./start.sh          # → http://localhost:3001
+            ./start.sh   # builds acestep.cpp, downloads models, starts UI → http://localhost:3001
             ```
 
-            > **GPU note:** bundles are built CPU-only.  
-            > Replace `bin/acestep-generate` with a CUDA/Metal build for GPU acceleration.
+            On Windows: extract the `.zip` and run `start.bat`.
 
             ### What's Changed
             ${{ steps.changelog.outputs.log }}
diff --git a/build.bat b/build.bat
new file mode 100644
index 0000000..9467983
--- /dev/null
+++ b/build.bat
@@ -0,0 +1,143 @@
+@echo off
+:: Build acestep.cpp with hardware-accelerated GPU support on Windows.
+:: Automatically detects CUDA (NVIDIA) and Vulkan.
+:: Called automatically by start.bat on first launch — or run manually to rebuild.
+::
+:: Usage: build.bat [options]
+::   --cuda      force CUDA build
+::   --vulkan    force Vulkan build
+::   --cpu       CPU-only build (disable GPU auto-detection)
+::   --src DIR   acestep.cpp source directory (default: .\acestep.cpp)
+::   --bin DIR   directory to install binaries into (default: .\bin)
+::   --repo URL  override the git repository to clone
+
+setlocal EnableDelayedExpansion
+
+set DIR=%~dp0
+set SRC_DIR=%DIR%acestep.cpp
+set BIN_DIR=%DIR%bin
+set BUILD_DIR=%SRC_DIR%\build
+if "%ACESTEP_CPP_REPO%"=="" set ACESTEP_CPP_REPO=https://github.com/audiohacking/acestep.cpp.git
+
+set FORCE_FLAGS=
+set CPU_ONLY=0
+
+:: Parse arguments
+:parse
+if "%~1"=="" goto :done_parse
+if /I "%~1"=="--cuda"   ( set FORCE_FLAGS=-DGGML_CUDA=ON   & shift & goto :parse )
+if /I "%~1"=="--vulkan" ( set FORCE_FLAGS=-DGGML_VULKAN=ON & shift & goto :parse )
+if /I "%~1"=="--cpu"    ( set CPU_ONLY=1                   & shift & goto :parse )
+if /I "%~1"=="--src"    ( set SRC_DIR=%~2 & set BUILD_DIR=%~2\build & shift & shift & goto :parse )
+if /I "%~1"=="--bin"    ( set BIN_DIR=%~2                  & shift & shift & goto :parse )
+if /I "%~1"=="--repo"   ( set ACESTEP_CPP_REPO=%~2         & shift & shift & goto :parse )
+echo Unknown option: %~1 & exit /b 1
+:done_parse
+
+echo ========================================
+echo   Building acestep.cpp
+echo ========================================
+echo.
+
+:: Check cmake
+where cmake >nul 2>&1
+if %ERRORLEVEL% NEQ 0 (
+    echo Error: cmake is required.
+    echo.
+    echo   Install from https://cmake.org/download/
+    echo   Or via winget:  winget install Kitware.CMake
+    echo   Or via choco:   choco install cmake
+    exit /b 1
+)
+
+:: Check git
+where git >nul 2>&1
+if %ERRORLEVEL% NEQ 0 (
+    echo Error: git is required.
+    echo.
+    echo   Install from https://git-scm.com/download/win
+    echo   Or via winget:  winget install Git.Git
+    exit /b 1
+)
+
+:: Clone if source is not present
+if not exist "%SRC_DIR%\.git" (
+    echo Cloning acestep.cpp from %ACESTEP_CPP_REPO% ...
+    git clone --depth 1 "%ACESTEP_CPP_REPO%" "%SRC_DIR%"
+    echo.
+) else (
+    echo acestep.cpp source found at %SRC_DIR%
+)
+
+echo Initializing submodules...
+cd /d "%SRC_DIR%"
+git submodule update --init --recursive --depth 1
+cd /d "%DIR%"
+echo.
+
+:: Determine cmake flags
+set CMAKE_FLAGS=
+if %CPU_ONLY%==1 (
+    echo CPU-only build requested -- skipping GPU detection
+    goto :skip_detect
+)
+if not "!FORCE_FLAGS!"=="" (
+    set CMAKE_FLAGS=!FORCE_FLAGS!
+    echo Using forced cmake flags: !CMAKE_FLAGS!
+    goto :skip_detect
+)
+
+:: Auto-detect GPU
+where nvcc >nul 2>&1
+if %ERRORLEVEL% EQU 0 (
+    set CMAKE_FLAGS=!CMAKE_FLAGS! -DGGML_CUDA=ON
+    echo Detected: CUDA ^(NVIDIA GPU^)
+)
+
+where vulkaninfo >nul 2>&1
+if %ERRORLEVEL% EQU 0 (
+    set CMAKE_FLAGS=!CMAKE_FLAGS! -DGGML_VULKAN=ON
+    echo Detected: Vulkan GPU
+)
+
+if "!CMAKE_FLAGS!"=="" echo No GPU accelerator detected -- CPU-only build
+
+:skip_detect
+echo.
+echo cmake flags:!CMAKE_FLAGS!
+echo.
+
+:: Configure
+cmake -S "%SRC_DIR%" -B "%BUILD_DIR%" -DCMAKE_BUILD_TYPE=Release !CMAKE_FLAGS!
+if %ERRORLEVEL% NEQ 0 ( echo cmake configure failed & exit /b 1 )
+
+:: Build
+cmake --build "%BUILD_DIR%" --config Release --parallel
+if %ERRORLEVEL% NEQ 0 ( echo cmake build failed & exit /b 1 )
+
+:: Copy binaries
+if not exist "%BIN_DIR%" mkdir "%BIN_DIR%"
+set COPIED=0
+for %%N in (ace-qwen3 dit-vae neural-codec) do (
+    set FOUND=
+    for /r "%BUILD_DIR%" %%F in (%%N.exe) do (
+        if exist "%%F" if "!FOUND!"=="" set FOUND=%%F
+    )
+    if not "!FOUND!"=="" (
+        copy /Y "!FOUND!" "%BIN_DIR%\%%N.exe" >nul
+        echo [OK] %%N.exe -^> %BIN_DIR%\%%N.exe
+        set /a COPIED+=1
+    ) else (
+        echo [WARN] %%N.exe not found in build output
+    )
+)
+
+echo.
+if !COPIED! GTR 0 (
+    echo Build complete^^! !COPIED! binaries installed to %BIN_DIR%\
+) else (
+    echo Error: build completed but no binaries were found in %BUILD_DIR%\.
+    echo Check the cmake output above for errors.
+    exit /b 1
+)
+echo.
diff --git a/build.sh b/build.sh
new file mode 100755
index 0000000..26103d0
--- /dev/null
+++ b/build.sh
@@ -0,0 +1,151 @@
+#!/bin/bash
+# Build acestep.cpp with hardware-accelerated GPU support.
+# Automatically detects CUDA (NVIDIA), ROCm/HIP (AMD), Vulkan, and Metal (macOS).
+# Called automatically by start.sh on first launch — or run manually to rebuild.
+#
+# Usage: ./build.sh [options]
+#   --src DIR      acestep.cpp source directory (default: ./acestep.cpp)
+#   --bin DIR      directory to install binaries into (default: ./bin)
+#   --cuda         force CUDA build
+#   --rocm         force ROCm/HIP build
+#   --vulkan       force Vulkan build
+#   --cpu          CPU-only build (disable GPU auto-detection)
+#   --repo URL     override the git repository to clone
+
+set -e
+
+DIR="$(cd "$(dirname "$0")" && pwd)"
+SRC_DIR="${ACESTEP_CPP_SRC:-$DIR/acestep.cpp}"
+BIN_DIR="${ACESTEP_BIN_DIR:-$DIR/bin}"
+BUILD_DIR="$SRC_DIR/build"
+REPO="${ACESTEP_CPP_REPO:-https://github.com/audiohacking/acestep.cpp.git}"
+FORCE_FLAGS=""
+CPU_ONLY=0
+
+# ── Parse arguments ───────────────────────────────────────────────────────────
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --src)    SRC_DIR="$2"; BUILD_DIR="$SRC_DIR/build"; shift ;;
+        --bin)    BIN_DIR="$2"; shift ;;
+        --repo)   REPO="$2"; shift ;;
+        --cuda)   FORCE_FLAGS="-DGGML_CUDA=ON" ;;
+        --rocm)   FORCE_FLAGS="-DGGML_HIP=ON" ;;
+        --vulkan) FORCE_FLAGS="-DGGML_VULKAN=ON" ;;
+        --cpu)    CPU_ONLY=1 ;;
+        *) echo "Unknown option: $1" >&2; exit 1 ;;
+    esac
+    shift
+done
+
+echo "========================================"
+echo "  Building acestep.cpp"
+echo "========================================"
+echo ""
+
+# ── Check required tools ──────────────────────────────────────────────────────
+missing=""
+command -v cmake &>/dev/null || missing="$missing cmake"
+command -v git   &>/dev/null || missing="$missing git"
+if [ -n "$missing" ]; then
+    echo "Error: the following required tools are missing:$missing"
+    echo ""
+    if [[ "$(uname)" == "Darwin" ]]; then
+        echo "  Install with: brew install$missing"
+    else
+        echo "  Install with: sudo apt-get install$missing  (Debian/Ubuntu)"
+        echo "           OR:  sudo dnf install$missing      (Fedora/RHEL)"
+    fi
+    exit 1
+fi
+
+# ── Clone or update acestep.cpp ───────────────────────────────────────────────
+if [ ! -d "$SRC_DIR/.git" ]; then
+    echo "Cloning acestep.cpp from $REPO ..."
+    git clone --depth 1 "$REPO" "$SRC_DIR"
+    echo ""
+else
+    echo "acestep.cpp source found at $SRC_DIR"
+fi
+
+echo "Initializing submodules..."
+cd "$SRC_DIR"
+git submodule update --init --recursive --depth 1
+cd "$DIR"
+echo ""
+
+# ── Detect hardware and set cmake flags ───────────────────────────────────────
+if [ "$CPU_ONLY" -eq 1 ]; then
+    CMAKE_FLAGS=""
+    echo "CPU-only build requested — skipping GPU detection"
+elif [ -n "$FORCE_FLAGS" ]; then
+    CMAKE_FLAGS="$FORCE_FLAGS"
+    echo "Using forced cmake flags: $CMAKE_FLAGS"
+elif [[ "$(uname)" == "Darwin" ]]; then
+    CMAKE_FLAGS=""
+    echo "Platform: macOS — Metal + Accelerate BLAS auto-detected by cmake"
+else
+    CMAKE_FLAGS=""
+
+    # CUDA (NVIDIA)
+    if command -v nvcc &>/dev/null || [ -d /usr/local/cuda ] || [ -d /usr/cuda ]; then
+        CMAKE_FLAGS="$CMAKE_FLAGS -DGGML_CUDA=ON"
+        echo "Detected: CUDA (NVIDIA GPU)"
+    fi
+
+    # ROCm/HIP (AMD)
+    if command -v hipcc &>/dev/null || [ -d /opt/rocm ]; then
+        CMAKE_FLAGS="$CMAKE_FLAGS -DGGML_HIP=ON"
+        echo "Detected: ROCm/HIP (AMD GPU)"
+    fi
+
+    # Vulkan
+    if pkg-config --exists vulkan 2>/dev/null || command -v vulkaninfo &>/dev/null; then
+        CMAKE_FLAGS="$CMAKE_FLAGS -DGGML_VULKAN=ON"
+        echo "Detected: Vulkan GPU"
+    fi
+
+    # OpenBLAS (CPU math acceleration — can coexist with GPU backends)
+    if pkg-config --exists openblas 2>/dev/null; then
+        CMAKE_FLAGS="$CMAKE_FLAGS -DGGML_BLAS=ON"
+        echo "Detected: OpenBLAS"
+    fi
+
+    [ -z "$CMAKE_FLAGS" ] && echo "No GPU accelerator detected — CPU-only build"
+fi
+
+echo ""
+echo "cmake flags: ${CMAKE_FLAGS:-<none>}"
+echo ""
+
+# ── Configure ─────────────────────────────────────────────────────────────────
+# shellcheck disable=SC2086
+cmake -S "$SRC_DIR" -B "$BUILD_DIR" \
+    -DCMAKE_BUILD_TYPE=Release \
+    $CMAKE_FLAGS
+
+# ── Build ─────────────────────────────────────────────────────────────────────
+cmake --build "$BUILD_DIR" --parallel
+
+# ── Copy binaries to bin/ ─────────────────────────────────────────────────────
+mkdir -p "$BIN_DIR"
+copied=0
+for name in ace-qwen3 dit-vae neural-codec; do
+    found=$(find "$BUILD_DIR" -name "$name" -type f 2>/dev/null | head -1)
+    if [ -n "$found" ]; then
+        cp "$found" "$BIN_DIR/$name"
+        chmod +x "$BIN_DIR/$name"
+        echo "✅ $name → $BIN_DIR/$name"
+        copied=$((copied + 1))
+    else
+        echo "⚠️  $name not found in build output — binary may have a different name"
+    fi
+done
+
+echo ""
+if [ "$copied" -gt 0 ]; then
+    echo "Build complete! $copied binaries installed to $BIN_DIR/"
+else
+    echo "Error: build completed but no binaries were found in $BUILD_DIR/."
+    echo "Check the cmake output above for errors."
+    exit 1
+fi
diff --git a/setup.sh b/setup.sh
index b7487e6..cc73dd5 100755
--- a/setup.sh
+++ b/setup.sh
@@ -28,19 +28,25 @@ if ! command -v curl &>/dev/null && ! command -v wget &>/dev/null; then
   echo "  sudo apt install curl  OR  brew install curl"
 fi
 
-# ── Check for acestep.cpp binary ─────────────────────────────────────────────
-if [ -n "${ACESTEP_BIN:-}" ] && [ -x "$ACESTEP_BIN" ]; then
+# ── Build acestep.cpp if binaries are not present ────────────────────────────
+ACESTEP_BIN_DIR="${ACESTEP_BIN_DIR:-bin}"
+if [ -x "$ACESTEP_BIN_DIR/ace-qwen3" ] && [ -x "$ACESTEP_BIN_DIR/dit-vae" ]; then
+  echo "acestep.cpp binaries: $ACESTEP_BIN_DIR/ ✓"
+elif [ -n "${ACESTEP_BIN:-}" ] && [ -x "$ACESTEP_BIN" ]; then
   echo "acestep-generate: $ACESTEP_BIN ✓"
 else
   echo ""
-  echo "Note: ACESTEP_BIN is not set."
-  echo "  Build acestep.cpp and set ACESTEP_BIN before generating music:"
-  echo ""
-  echo "    git clone https://github.com/audiohacking/acestep.cpp"
-  echo "    cmake -S acestep.cpp -B acestep.cpp/build -DCMAKE_BUILD_TYPE=Release"
-  echo "    cmake --build acestep.cpp/build --parallel"
-  echo "    export ACESTEP_BIN=\$(pwd)/acestep.cpp/build/bin/acestep-generate"
+  echo "Building acestep.cpp for your hardware (detects GPU automatically)..."
   echo ""
+  bash build.sh || {
+    echo ""
+    echo "  Automatic build failed. You can build manually:"
+    echo "    git clone https://github.com/audiohacking/acestep.cpp"
+    echo "    cmake -S acestep.cpp -B acestep.cpp/build -DCMAKE_BUILD_TYPE=Release"
+    echo "    cmake --build acestep.cpp/build --parallel"
+    echo "  Then set ACESTEP_BIN_DIR=$(pwd)/bin in .env"
+    echo ""
+  }
 fi
 
 # ── Download GGUF models ─────────────────────────────────────────────────────