Running local AI models on Microsoft’s Surface Laptop Ultra: What we found
The New Stack Frederic Lardinois ● Covered by 8 sources
Microsoft’s new Surface Laptop Ultra runs local AI models well. It even keeps up with a Mac Studio on some workloads, but battery life drops fast.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft is trying to sell the Surface Laptop Ultra as the Windows answer for creative pros and AI developers. On paper, it has the right kind of muscle: Nvidia’s N1X-based RTX Spark platform, a Blackwell RTX GPU with up to 6,144 cores, a CPU with up to 20 cores, and up to 128GB of unified memory. In practice, the pitch is simple enough: a laptop that can do serious local AI work without feeling like a downgrade.
That part mostly holds up. The unit tested here was the high-end $4,300 configuration with 64GB of RAM and a 1TB SSD, and over a week of use it handled the usual local-model tools without drama. Unsloth Studio, Ollama, Hermes Agent and the rest all ran as expected. In some cases, the Blackwell GPU even felt a little faster than a Mac Studio with an M4 Max, which is not a trivial result if your day involves shipping and testing models rather than talking about them.
The speed depends heavily on the model. Nemotron 3.5 Bolt, a heavily quantized 9B model, hit about 56 to 60 tokens per second and occasionally more. Dense models were another story: Qwen 3.8 27B at 8-bit quantization never got past 8 or 9 tokens per second, which is basically the memory bandwidth tax showing up in real time. Drop that model to 4-bit quantization and it climbed back to roughly 25 to 28 tokens per second. Gemma 4 26B-A4B, at 8-bit, ran at around 40 tokens per second.
The big catch is power. Those numbers came from the 140W charger, and once the laptop was unplugged the AI workload slowdown was obvious. The 4-bit Qwen model fell to around 18 tokens per second, and Nemotron 3.5 Bolt dropped to about 42 to 45 tokens per second. Microsoft says the machine can keep 99.6% of its performance on battery, but that did not match what was seen here. For constant local AI work, a few hours of battery life sounds more realistic than a workday.
There are trade-offs elsewhere too. Sixty-four gigabytes of RAM leaves the machine in a strange middle ground: enough headroom for local models and the rest of the desktop, but not so much that the available model sizes suddenly become ideal. Going to 128GB pushes the price to $5,900. Heat and fan noise are part of the deal, though not constant. The laptop is quiet most of the time, then gets noticeably loud under heavier AI loads, especially when a coding agent keeps chewing on a query.
Still, this is a more convincing Windows ARM laptop than people might expect. The ports are generous, the SSD is replaceable, and the 15-inch 3:2 touchscreen looks good even if it is reflective. Windows on ARM also feels more mature than it did a year ago, and the machine can even game surprisingly well: Cyberpunk 2077 ran at around 60fps in 2K with ray tracing on, while The Witcher 3 remaster hit 60fps in 4K with ray tracing off. That may not be the main act, but it says something about where Microsoft wants this platform to go.
My take — AI-written commentary, not fact-checked reporting
Microsoft is finally making a Windows machine that doesn’t apologize for existing next to a MacBook Pro. The annoying part is the usual one: once the hardware gets interesting, the price and the fan noise arrive to remind everyone that “portable AI workstation” is still a compromise dressed in good aluminum. Open tools are clearly getting better on Windows ARM, and that’s the real story here, not the marketing gloss.
Read more about this at: The New Stack
