{
  "id": 14950828,
  "title": "Running local AI models on Microsoft’s Surface Laptop Ultra: What we found",
  "url": "https://urgent.news/2026/10/16/running-local-ai-models-on-microsofts-surface-laptop-ultra-what-we",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-16T12:00:00.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/surface-ultra-local-ai/"
  },
  "original_language": "en",
  "account": "Microsoft is positioning the Surface Laptop Ultra as a competitive alternative to Apple's MacBook Pro, especially for AI developers. The laptop is equipped with Nvidia's N1X-based RTX Spark platform, which features a Blackwell RTX GPU with up to 6,144 cores, a CPU with up to 20 cores, up to 128GB of unified memory, and a price that is lower than Apple's offerings. This makes it an attractive option for those who want to run local AI models alongside powerful video editing and gaming capabilities. This article focuses on testing the laptop's performance for AI workloads, leaving gaming benchmarks and creative use cases to others. The laptop was tested in its high-end configuration with a $4,300 GPU and CPU setup, 64GB of RAM, and a 1TB SSD. It was used for a week, including while traveling. Typically, the user relies on a Mac Studio with the M4 Max chip at home, which can run most local models (provided they fit within the 48 GB of RAM). The user installed common tools like Unsloth Studio, Ollama, and Hermes Agent on the laptop. The Blackwell GPU generally performed faster than the Mac Studio for local inference with the right models. For example, the Nemotron 3.5 Bolt model responded to queries at a rate of 56 to 60 tokens per second, making it a fast option for Hermes Agent or OpenCode. However, denser models like Qwen 3.8 27B at 8-bit quantization only achieved 8 to 9 tokens per second, a rate that is usually limited due to memory bandwidth. Using 8-bit Gemma 4 26B-A4B, the mixture-of-experts technique boosted performance to around 40 tokens per second. The Surface Laptop Ultra generally outperformed the M5 Max in MacBook Pro AI use cases. Performance dropped when running on battery, with models like 4-bit Qwen 3.8 dropping from 25-28 tokens per second to around 18 tokens per second. The laptop is expected to retain around 20% to 25% of its performance when on battery. The user experienced a 20% to 25% drop in performance when running AI workloads on battery, similar to gaming performance. The laptop's RAM may be excessive for smaller models, but it provides ample headroom for KV cache and other applications. However, 128GB of RAM increases the price to $5,900. The laptop generates heat and noise, with dual fans that are noticeable during intensive use. Despite this, the fans remain relatively quiet during most day-to-day tasks. The touchpad is large and features haptic feedback, enhancing the editing experience. The speakers deliver excellent sound quality, with wide stereo field. The magnetic USB-C charging connector resembles other Surface devices, but there is concern about losing the cable. Microsoft did not position the Surface Laptop Ultra as a gaming PC, but some users may attempt to play games on the device, especially with its powerful GPU.",
  "summary": "Microsoft is positioning the new Surface Laptop Ultra as an alternative to Apple’s MacBook Pros for creatives and developers — The post Running local AI models on Microsoft’s Surface Laptop Ultra: What we found appeared first on The New Stack .",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Wired",
        "title": "Microsoft Surface Laptop Ultra Review: The First Real AI PC",
        "url": "https://urgent.news/2026/10/16/microsoft-surface-laptop-ultra-review-the-first-real-ai-pc",
        "published": "2026-10-16T13:00:00.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}