Logan Iyer, Microsoft’s corporate vice president for Windows platform and developer, gave a name on Sept. 4, 2026, to the developer-optimized Windows 11 experience first shown at Build in June. Project Zenith would arrive with coding tools and settings already installed, first on PCs built around AMD’s Ryzen AI Halo processors.

A Lenovo machine in the category arrives in November 2026 with a starting price of $3,699.

Project Zenith pairs a preconfigured Windows 11 installation with a hardware class defined on Sept. 4, 2026, by at least 64 GB of unified memory and 250 GB/s of memory bandwidth. The pitch is that developers can run large AI models on their desks without paying for every token sent to a cloud service.

What Changed

AI-generated summary, reviewed by an editor. More on our AI guidelines.

The hardware floor

Unified memory gives the processor and graphics hardware access to the same pool of RAM. That lets a machine hold models too large for the 24 GB or 32 GB of video memory common in high-end consumer graphics cards, though fitting a model says little about how quickly it will respond.

Lenovo’s ThinkCentre X Ultra shows what the category costs. The 1.6-liter desktop, announced on Sept. 3, 2026, for shipment that November, pairs an AMD Ryzen AI Max+ PRO 495 processor with up to 128 GB of unified memory and an NPU rated at about 55 TOPS. Up to four units can be clustered, although Lenovo has not disclosed the effective bandwidth or performance of that arrangement.

Nvidia’s competing DGX Spark also carries 128 GB of unified memory. It launched on Oct. 15, 2025, at $3,999, then rose 18 percent to $4,699 on Feb. 23, 2026, a change Nvidia tied to worldwide memory-supply constraints. Its 273 GB/s bandwidth is only modestly above Microsoft’s Zenith threshold.

AMD showed its Ryzen AI Halo mini PC on stage on Sept. 4, 2026, and the announcement included no price. The $3,699 starting point belongs to Lenovo’s separate ThinkCentre configuration for November 2026.

What the 30B claim means

“On these devices, developers can run 30B+ parameter models locally and unmetered,” Iyer said in the Sept. 4, 2026, announcement. That describes capacity. Speed depends on how much of a model the computer must read from memory to produce each token.

A dense model activates all its parameters for every token. A mixture-of-experts model keeps the full model in memory but calls only a smaller group of parameters for each step, like a workshop sending a job to a few specialists instead of gathering the entire staff.

That distinction produces large differences on Ryzen AI Max+ 395 hardware with roughly 256 GB/s of theoretical bandwidth. Community benchmark threads cited in a compilation updated Aug. 29, 2026 put Qwen3-30B-A3B, a mixture-of-experts model with about 3 billion active parameters per token, at roughly 70 to 100 tokens per second. The broader reported envelope put a dense 70-billion-parameter model at 4-bit at about 5 tokens per second. ServeTheHome measured GPT-OSS 120B, another mixture-of-experts model, at about 31 tokens per second, while the broader reported range for dense models in the 7-billion to 13-billion range was roughly 30 to 45.

Those figures combine community benchmarks, ServeTheHome’s measurement and broader reported performance ranges on Ryzen AI Max+ 395 systems, not results from Microsoft or a Project Zenith device. The Sept. 4 announcement carried no tokens-per-second figures and named no specific model behind the “30B+” claim.

The Windows configuration

The software layer installs Visual Studio Code, GitHub Copilot, PowerToys, WinAppCLI and Windows Dev Skills. Its September 2026 baseline includes Python 3.14 or newer, Node 24 or newer, WSL 2 with Ubuntu and .NET 10. File Explorer shows extensions, hidden files and full paths; long-path support is enabled; Start menu tips and account notifications are turned off.

Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.

The same opinionated setup is publicly available through Microsoft’s Windows Developer Configuration repository. It can apply the tools and settings to an existing Windows 11 machine through winget, with a required restart when WSL is enabled.

Microsoft says Project Zenith devices benefit from the platform work for AI agents it showed at Build 2026 from day one. The company says that work combines OS-enforced identity, containment through Microsoft Execution Containers (MXC) and enterprise-grade manageability.

Sean Endicott, a Windows Central writer, said, “What would be considered a useful preinstalled app for developers would be viewed as bloat for general users.”

The configuration problem

Project Zenith arrives during Pavan Davuluri’s effort to repair confidence in Windows 11. On March 20, 2026, the Windows chief outlined fewer automatic restarts, less Copilot integration in Snipping Tool, Photos and Notepad, and a single monthly reboot. “What came through was the voice of people who care deeply about Windows and want it to be better,” Davuluri said.

Paul Thurrott, who writes at Thurrott.com, tested the public configuration that underpins Zenith and called the project “a curious miscalculation.” Developers already have individual setups, he said, and will spend time reversing defaults that do not fit their work. He proposed a more useful Windows Backup that could restore a developer’s own configuration on any PC.

The test ended with a full wipe. “I had to wipe the PC I tried this on, it was maddening,” Thurrott said.

Frequently Asked Questions

What is Project Zenith?

A preconfigured Windows 11 installation for developer-class PCs, named by Microsoft on Sept. 4, 2026. It arrives with coding tools and settings already installed, first on PCs built around AMD's Ryzen AI Halo processors.

What hardware qualifies as a Project Zenith device?

Microsoft set the floor on Sept. 4, 2026 at a minimum of 64 GB of unified memory and 250 GB/s of memory bandwidth.

What does a Project Zenith machine cost?

Lenovo's ThinkCentre X Ultra starts at $3,699 and ships in November 2026. AMD showed its Ryzen AI Halo mini PC on Sept. 4, 2026 and the announcement included no price. For comparison, Nvidia's DGX Spark rose from $3,999 to $4,699 in February 2026.

Can these devices actually run 30B parameter models?

They can hold them. How fast they run depends on architecture. Community benchmarks put a 30B mixture-of-experts model at roughly 70 to 100 tokens per second on comparable hardware, while a dense 70-billion-parameter model at 4-bit managed about 5. The Sept. 4 announcement carried no tokens-per-second figures.

Can developers get the Zenith setup without buying a new PC?

Yes. The same configuration is publicly available through Microsoft's Windows Developer Configuration repository, and it applies the tools and settings to an existing Windows 11 machine through winget, with a restart required when WSL is enabled.

AI-generated summary, reviewed by an editor. More on our AI guidelines.

Perplexity Ships Local AI Agent on Nvidia DGX Spark With No On-Device Token Cost
Perplexity launched Portable Computer on Tuesday, moving its Computer agent stack onto Nvidia hardware so files, models and most tool execution can stay on a user's machine. Work completed locally con
The Implicator
Meta Opens Muse Glimmer as Zuckerberg Presses U.S. AI Policy Shift
On Aug. 10, 2026, Meta released Muse Glimmer, a 30-billion-parameter model whose weights developers can download. It was built for agents that run on a Mac or PC. The release carried a policy argumen
The Implicator
Apple expands Xcode and Foundation Models for WWDC AI developers
Apple said Monday it expanded Xcode and the Foundation Models framework at its 2026 Worldwide Developers Conference (WWDC), giving developers image input, custom skills and server-side model execution
The Implicator
Analysis

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai