The useful distinction is that Zenith prepares the development environment; Halo supplies the hardware. It is not a replacement operating system or an AI model. Microsoft describes it as a tailored Windows experience that developers can continue to customize, rather than a prescribed workflow.
What Project Zenith changes in Windows 11
Microsoft announced Zenith on September 4, 2026, naming Halo as its first platform and promising additional OEM and silicon-partner devices in the following months. Its stated hardware baseline includes at least 64 GB of unified memory and 250 GB/s of memory bandwidth.
The configuration tackles familiar setup chores:
- Windows Terminal and Visual Studio Code are pinned to the taskbar.
- File Explorer displays extensions, hidden files and full paths.
- Long-path support is enabled.
- Recent-file suggestions, sync-provider tips and certain Start notifications are disabled.
AMD lists Git, GitHub CLI, GitHub Copilot CLI, Azure CLI, Python, Node.js,.NET, PowerShell 7 and Windows Subsystem for Linux among the installed tools. It also adds Ryzen AI Developer Center, validated software configurations and guided AI Playbooks.
The practical appeal is less ceremony between unpacking a PC and opening a project. That is a useful starting point—not a claim that every project arrives configured.
The memory headline needs a performance footnote
AMD says Halo offers up to 192 GB of unified memory, with up to 160 GB available to graphics and 273 GB/s of bandwidth. The company associates that capacity with models containing up to 300 billion parameters, longer contexts and multiple-model workflows.
Capacity is not a speed guarantee. AMD’s own getting-started documentation distinguishes memory capacity from generation performance and notes bandwidth limits for large models. Its optional memory-allocation guidance also recommends leaving roughly 20% of system RAM for the operating system.
The editorial takeaway: evaluate a particular model and workflow, not just the largest parameter count on the box.
Preinstalled software does not mean preinstalled models
An important operational detail appears in AMD’s Halo playbook: the Windows image includes Lemonade Server, ComfyUI and ROCm-enabled PyTorch, but model files are not included. These are AMD’s AI components alongside Zenith, not a universal promise about every Zenith device.
For owners of that Windows Halo image, AMD documents this initial sequence:
- Install pending updates through Ryzen AI Developer Center and reboot.
- Launch Lemonade Server from Start.
- Open its interface at
[url]http://localhost:13305[/url]. - Choose and download a model.
- Use the exact model identifier when connecting a client to
[url]http://localhost:13305/v1[/url].
Success means the local server is accessible and a downloaded model is available to the client. An empty model list is not necessarily a broken installation: the image starts without model downloads. A rejected identifier warrants checking its spelling against Lemonade’s reported name.
Existing Halo owners have another route
AMD says owners already running Windows on Ryzen AI Halo can configure the environment using Microsoft’s Windows Developer Config project. A new purchase is therefore not required solely to obtain the developer configuration.
That is the strongest framing for this launch: a more deliberate Windows developer baseline, paired with AMD’s local-AI software and high-memory hardware. The promise worth judging is not “every enormous model, effortlessly.” It is whether a prepared machine gets developers to a working, appropriate model with fewer setup detours.