Surface Ultra AI benchmarks fall 20 to 25 percent on battery

A laptop that keeps almost all its speed on battery is a laptop you can take to a coffee shop and run a model on. The New Stack's Frederic Lardinois spent a week with Microsoft's Surface Laptop Ultra, in the $4,300 configuration with 64GB of RAM, running four local AI models on it, and he found the plug matters more than Microsoft said it would.

Microsoft claims the Ultra holds 99.6 percent of its performance on battery, Lardinois reports. His own measurements put the drop under AI load at 20 to 25 percent. A 4-bit Qwen 3.8 that ran at 25 to 28 tokens per second plugged in fell to around 18 tokens per second on battery, which is a drop closer to 35 percent for that model. The 9B-active Nemotron 3.5 Bolt went from 56 to 60 tokens per second plugged in to 42 to 45 on battery.

The models he tested

The Ultra is built on Nvidia's N1X RTX Spark platform with a Blackwell GPU, up to 6,144 cores, up to a 20-core CPU, and up to 128GB of unified memory, Lardinois writes. He used Unsloth Studio, Ollama, and Hermes Agent, the same stack he runs on a Mac Studio with an M4 Max and 48GB of RAM at home.

The 8-bit Gemma 4 26B with a mixture-of-experts architecture reached about 40 tokens per second on the Ultra, a speed Lardinois says is usable for a coding agent or a chat workflow. The 8-bit dense Qwen 3.8 27B topped out at 8 or 9 tokens per second, which he writes is too slow to use, and he attributes the result to memory bandwidth, not compute. Bringing that model down to 4-bit was what moved it to the 25 to 28 range. He writes that in general the Blackwell GPU felt faster than his M4 Max.

The gaps that stay with you

Battery life is the gap under battery performance. Microsoft says 12 hours of active web use. Lardinois did not run a formal test, but based on his observation, running local AI workloads continuously would give a few hours of battery, not twelve.

The second gap is memory tier. The 64GB machine covers today's 27B to 30B models and leaves room for the KV cache, Lardinois writes, but there are few models in the 32GB to 48GB range the Ultra could also handle, which puts the extra memory in an awkward middle ground. Moving to 128GB raises the price to $5,900. The third gap is noise. Under sustained inference the dual fans are not whisper-quiet, and he notes the matching Surface RTX Spark Dev Box has, in his words, a thousand air vents to push the same N1X platform harder.

Lardinois also compared the Ultra to a model he does not have on hand. He writes that Tom's Hardware reports the Ultra generally outperforms an M5 Max in a MacBook Pro on AI workloads, so his M4 Max comparison is one generation back on the Apple side.

The pattern this fits is a vendor benchmark measured on a workload that happens to use the parts of the machine the vendor tuned for battery, and a user workload that uses the parts they did not. A developer picking a laptop to run local models should ask vendors for the AI numbers under battery, not the aggregate.

Source

The New Stack, Frederic Lardinois, Running local AI models on Microsoft's Surface Laptop Ultra: What we found.