# Windows x64 beta The first Windows beta uses the CPU for built-in answer generation, including on computers with NVIDIA GPUs. Use **Auto** or **CPU** in the local model settings. The model runs offline after you download it in Graphite. [Find your beta](https://github.com/mbagagli/graphite/releases) and look for the `Graphite__x64-setup.exe` asset on a published release. Check its SHA-256 against the release's `SHA256SUMS.txt` before running the unsigned NSIS installer. Setup may need internet access if Microsoft Edge WebView2 Runtime is missing. Published releases list their available platform assets. ## Existing starting point The `Native compatibility` workflow checks Windows x64 on `windows-2025`, independently of Linux/macOS releases. It checks packaging configuration before linking the development application or building the installer. The manual `windows_installer` input uses NSIS and the existing `windows-x64` inspection and installed-runtime checks. The release workflow builds its Windows x64 candidate on `windows-2025`, then runs the same inspection and installed lifecycle before including it in the release draft. See the [release workflow](BUILD.md). ## First Windows installer experiment Start with native Windows x64 and CPU answer generation. The locked Windows ONNX dependency enables DirectML; inspect its actual runtime dependencies even when answer generation uses CPU. Before tagging, run **Unsigned package and installed runtime smoke** on the reviewed branch. It checks Linux, macOS, and Windows and uploads each inspected installer with its checksum. Check the Windows lifecycle result before using the artifact for desktop acceptance. For a Windows-only run, use **Native compatibility** with `windows_installer` enabled once that workflow is registered on the repository's default branch. It uploads the inspected installer and checksums as `graphite-windows-x64-unsigned-experiment` for seven days. Use a disposable Windows host with the repository's pinned Node/npm and Rust versions, Microsoft C++ Build Tools, WebView2, CMake and LLVM/libclang. Use an x64 MSVC developer shell. For prerequisites and NSIS/WebView2 options, see the official [Tauri prerequisites](https://v2.tauri.app/start/prerequisites/) and [Windows installer guide](https://v2.tauri.app/distribute/windows-installer/). In PowerShell, from the repository root: ```powershell npm ci $clang = Get-Command clang.exe -ErrorAction Stop $env:LIBCLANG_PATH = Split-Path $clang.Source npm run package:check npm run tauri -- build --ci --bundles nsis -- --locked npm run package:smoke -- windows-x64 ``` These are proposed validation commands, not a record of a successful run. The native build uses the Windows platform configuration automatically. Omitting an explicit target keeps output in the package inspector's expected directory. Inspection requires 7-Zip and LLVM's `llvm-objdump` (or a PE-capable `objdump`) on PATH. Do not use `npm run package:build` for this experiment: its current platform guard rejects Windows. ## Beta acceptance 1. Run a hosted packaging check. Retain the inspected NSIS installer and checksums for testing. 2. Check the installer inspection results. The inspector checks the x64 executable's DLL imports and rejects separately distributed inference runtimes and CUDA DLLs. Validate ONNX and its system dependencies on a clean host. Check that setup can provision WebView2 when needed. 3. Check that the hosted installed-runtime step passed. It checks first launch, state isolation, uninstall/reinstall and state preservation. 4. Test the exact installer on a clean Windows desktop: graph rendering, PDF reading/export, read-only Zotero access, explicit model download, semantic indexing, CPU answers and reopening saved research. Check paths containing spaces/non-ASCII characters and both existing storage layouts. Test upgrade from an earlier candidate as well as reinstalling one build. 5. Publish the release draft after the exact Windows installer completes acceptance. Check the asset and checksum on the public Releases page. After the CPU installer has a reproducible install, application and upgrade baseline, test a CUDA build on a Windows NVIDIA host. Check that the package includes the necessary CUDA redistributable DLLs and that Graphite still launches and answers on a machine without an NVIDIA driver. Validate the two paths before making CUDA part of the standard Windows installer.