On November 20, LocalAI released a new version that lets the server and AI model runtime run directly on Windows. Developers can download a dedicated executable. GPU deployment users should check the changed image build defaults.
LocalAI v5.0.1 expands runtime support and revises model replacement. Existing files are now preserved until new model files pass validation, keeping the previous files in place while replacements download.
1. Windows support covers both the server and model runtime
The release notes announce native Windows support for the server and the windows/amd64 llama.cpp backend. The backend performs model inference. Direct Windows support for both components is the key change.
Release assets include a Windows server executable. Developers who want to run the server directly can download local-ai-v5.0.1-windows-amd64.exe from the release page.
The same page separately offers an installer and archive for the Windows launcher, which helps start the server. With separate server and launcher downloads, users should choose the asset for the component they need.
2. Default build settings change for GPU images
For GPU deployment, hipblas images now build with ROCm 10.1 by default. This changes the baseline for deployment images. Users building these images should first check whether their settings specify a version or use the default.
Related paths were also fixed. *_TENSILE_LIBPATH now points to the folder for the installed GPU. Those managing GPU image build workflows can review uses of this path alongside version settings.
The privacy filter add-on, an extension providing extra functionality, gains a ROCm-based hipblas image build. Beyond changing the base image build settings, this adds a GPU image deployment option for the privacy filter.
3. Model and add-on downloads get improvements
The model download fix delays replacing existing files until after validation. Instead of removing old files before checking the new ones, it keeps them until checks finish. This preserves existing files during model downloads.
Fetching the add-on index, a list used to browse available items, was also fixed. It now follows redirects from GitHub releases. This handles cases where the address changes while retrieving the add-on list.
The add-on catalog now attaches to build outputs without manually rerunning the job. Separate from the user-facing index retrieval fix, this reduces manual work when producing release artifacts.
4. More runtime options, different deployment checks
This release adds components that run directly on Windows and more GPU deployment images. ROCm image builders need to review default versions and GPU paths. Initial checks depend on whether users build images or download executables.
The release notes specify Windows support for the server and the corresponding 64-bit backend. They provide no performance measurements for the ROCm default change.
LocalAI v5.0.1 expands runtime and deployment options; adoption decisions should focus on the Windows components and the new GPU image build settings.
