Surface Laptop Ultra Tested for Local LLM Performance
October 19, 2026 · 2 min read
The Surface Laptop Ultra has been tested for its ability to run local LLM models, demonstrating functionality but facing issues with overheating. This testing was conducted by The New Stack and focused on the RTX Spark platform, which is part of NVIDIA's latest push into the PC market after the Tegra and Surface RT era. Various brands, including Microsoft's Surface Laptop Ultra, are now available in the United States with this new hardware.
The tests utilized a model equipped with 64GB of RAM, which is not the highest specification available at 128GB. The performance results were notable, with the new NVIDIA model, Nemotron 3.5 Bolt, achieving 56-60 tokens per second, while the Qwen 3.8 model achieved 8-9 tokens per second with 8-bit quantization and 25-28 tokens per second with 4-bit quantization. The Gemma 4 26B-A4B model reached 40 tokens per second, attributed to its mixture-of-experts (MoE) architecture.
However, when the device was unplugged, performance significantly decreased. For instance, the Qwen 3.8 with 4-bit quantization dropped to 18 tokens per second, and the Nemotron 3.5 Bolt fell to 42-45 tokens per second, reflecting an overall decrease of about 20% in efficiency.
Additionally, concerns were raised about the Surface Laptop Ultra’s temperature during use. The device's fans operated at full speed, producing noticeable noise while running LLM models, although it remained quiet during other tasks. For users intending to maximize RTX performance for model running, a better option might be the Surface RTX Spark Dev Box, which is designed with improved ventilation.
The availability of models that can run on 64GB of RAM is limited, mainly supporting sizes around 27B-30B. Users are advised to consider the 128GB RAM option for better compatibility, although this comes at a high price tag of 5,900 dollars, equivalent to approximately 200,000 baht.