Windows 11 has just posted a series of startling wins in local AI benchmarks against two Linux distributions on identical high-end hardware. In a 99-test gauntlet run by Phoronix, Microsoft’s OS dominated CPU-based Llama.cpp text generation and speech-to‑text tasks, yet the overall crown went to the performance‑tuned CachyOS.

The tests, published on July 20, pitted Windows 11 against Ubuntu 26.04 LTS and CachyOS on a Razer Blade 18 laptop equipped with an Intel Core Ultra 9 290HX Plus processor and Nvidia GeForce RTX 5090 Laptop GPU. The results flip some long‑held assumptions—while Linux remains the go‑to for video encoding and rendering, Windows suddenly looks like a serious player for anyone running AI workloads directly on the CPU.

The Benchmark Breakdown: Where Windows 11 Shone

The testbed was a single, top‑tier machine: Razer’s Blade 18 RZ09‑0582, shipping with Windows 11. Phoronix installed Ubuntu 26.04 LTS (Linux 7.0) and the rolling CachyOS (Linux 7.1) alongside. Drivers mattered—Windows and Ubuntu used Nvidia’s R595 branch, while CachyOS benefited from the newer R610 series.

Windows 11 delivered its cleanest victories in Llama.cpp when using the CPU OpenBLAS backend. It was the fastest in every measured text‑generation and prompt‑processing test across three model families: Qwen, GPT‑OSS, and GLM. For anyone running quantized local models without a GPU, those numbers translate to snappier completions and lower latency.

When the Vulkan backend was engaged, things got muddier. CachyOS led several text‑generation tests, but Windows 11 still pulled ahead on prompt processing for two models. Phoronix called Vulkan performance “similar across Windows and Linux” overall.

Speech‑to‑text tasks brought another Windows surprise. Using Whisper.cpp‑based Whisperfile with Small and Medium models, Windows 11 outran both Linux rivals by what Phoronix described as an “unexpectedly large” gap. That result is specific to this hardware‑driver‑software stack, but it signals that Windows may be a practical choice for local transcription workloads.

How We Got Here: AI’s Shift to the Local Machine

Not long ago, running AI locally meant wrestling with Linux drivers and Python environments. But tools like llama.cpp, which compiles into a single executable with almost no dependencies, have made CPU inference mainstream—and cross‑platform from day one. Whisper.cpp followed the same pattern, democratizing speech‑to‑text.

Microsoft has been quietly optimising for this new world. The DirectML API, oneAPI support, and tighter integration with the Windows Subsystem for Linux have all improved ONNX and inference runtimes. And while Nvidia’s CUDA ecosystem is deeply Linux‑native, the Windows driver stack has matured enough to level the playing field, especially for CPU‑bound models.

Phoronix’s test suite, however, didn’t just look at AI. It threw 99 different benchmarks at the machines—encoding, rendering, compilation, and more—to paint a complete picture. That’s where the headline numbers get complicated.

Beyond AI: The Full 99‑Test Picture

Across all 99 tests, CachyOS took first place 51% of the time. Windows 11 managed 38%, and Ubuntu trailed at 10%. Windows was dead last in roughly half the benchmarks. On the geometric‑mean aggregate, CachyOS clung to a narrow lead, with Windows just ahead of Ubuntu.

But those aggregate rankings need a huge asterisk. CachyOS is an explicitly performance‑tuned Arch‑based distribution, built with newer kernels, compilers, and in this case a newer Nvidia driver. The comparison shows what you get out of the box with each platform, not a controlled OS‑only variable.

In media workloads, Linux largely held its traditional advantage. Ubuntu and CachyOS were faster in OpenAPV encoding. SVT‑AV1 encoding “typically” performed best on Linux, with CachyOS taking most of those wins. CPU rendering and several GPU‑accelerated benchmarks also favoured Linux—except for Chaos V‑Ray’s Nvidia RTX test, where Windows 11 came out on top.

Video encoding was a split decision: Windows 11 won several 1080p Kvazaar H.265 tests, while CachyOS led the 4K slow and medium presets. For content creators, the message is that Linux still offers better performance in many open‑source encoding tools, but Windows isn’t a slouch for certain tasks.

What It Means for You

If you’re an AI hobbyist or developer running local LLMs on a Windows laptop, the Phoronix numbers are genuinely good news. For CPU‑only inference using llama.cpp—say, with a 7B parameter quantised model—you can expect competitive or even class‑leading speeds without dual‑booting. Speech‑to‑text via Whisperfile also looks surprisingly strong. These aren’t synthetic differences; a few hundred milliseconds shaved off each token can transform the feel of an interactive chat.

Power users and IT pros managing heterogeneous fleets should note that Windows now stands toe‑to‑toe with Linux for a set of AI workloads that used to be Linux’s exclusive territory. That might simplify deployment if your organisation is Windows‑first.

Home users eyeing a new laptop for creative work will still find Linux compelling for video encoding and rendering. If you spend your days in Blender, HandBrake, or AV1 workflows, the performance delta remains meaningful—especially with a tuned distro like CachyOS. But for general use, Windows 11’s raw speed in everyday programs is no longer embarrassed by the penguin.

Developers building cross‑platform AI apps should test on both OSes. The Vulkan backend results in particular suggest that driver maturity can tip the scales. Your mileage will vary depending on model architecture, batch size, and hardware configuration.

What to Do Now

No action is mandatory, but a few steps can help you make the most of these findings:

  • If you’re running local LLMs on Windows, stick with the CPU OpenBLAS backend unless you have a large GPU. Consider the Whisperfile tool for fast transcription. Keep your Nvidia drivers up to date—the R595 branch used in the test is already widely available.
  • If you dual‑boot, measure your own workloads. A quick Llama.cpp test with your favourite model can tell you which OS gives you lower tokens‑per‑second. The Phoronix config is a useful reference point, but your specific model quantisation level and prompt length will change the outcome.
  • If you’re building a video‑encoding workstation, Linux remains the safer bet. CachyOS or a well‑tuned Ubuntu install will likely give you faster encodes. Windows users doing occasional H.265 1080p work can stay put without a huge penalty.

No deadlines loom, but the landscape is shifting fast. As Nvidia releases new driver branches and Microsoft integrates more AI acceleration into Windows (think Copilot+ PC features and DirectML improvements), these numbers could look very different in six months.

The Road Ahead

Windows 11 has proven it can punch above its weight in local AI—a territory Linux has long owned. With Arm‑based Windows devices and dedicated NPUs entering the mix, the performance map will only get more complex. Phoronix’s test is a snapshot: one laptop, one driver set, one point in time. But it’s a snapshot that should make every Windows user who cares about AI pay attention. Keep an eye on upcoming driver releases and distro updates; the next round of benchmarks may tell an even more interesting story.

For now, the practical takeaway is simple: if you want to run a local language model on your Windows laptop, go ahead. You won’t be leaving performance on the table. And if you also need to encode a video later, well, you might just want to keep that Linux USB stick handy.