Ars Technica runs a 70-billion-parameter AI model on Surface, but finds software gaps

Ars Technica’s October 16 review puts a concrete result behind Microsoft’s local-AI laptop pitch: a 70-billion-parameter model running on the Surface Laptop Ultra. Yet its tests also exposed a software gap. Some AI tools could use only the main processor, leaving the Nvidia graphics hardware unavailable for those workloads.

The hardware takes a different route from chips that use a dedicated neural processing unit, or NPU, for AI work. According to Nvidia, as described by Ars, the RTX Spark N1X combines a Grace central processor with a Blackwell graphics processor. It has no dedicated NPU. Instead, it relies on the graphics processor and unified memory—a pool shared by the processors—to handle AI workloads.

Reviewer Nick Indge tested a configuration with a 20-core CPU, a 6,144-core graphics processor and 64GB of unified memory. That machine lists for $4,299.99. The $2,599.99 starting price belongs to a different configuration with 24GB of memory, not the system behind his large-model result.

When software leaves the graphics processor idle

Microsoft recommended that reviewers try AI Dev Gallery, its public-preview app for testing on-device AI. The app lets users pull models from Hugging Face. In Indge’s testing, however, access to the laptop’s AI hardware depended on which model or tool he chose:

• Conversational models: All available Phi and Mistral options in AI Dev Gallery were CPU-only, with no route to GPU or NPU acceleration. That restriction did not apply across the entire app: some image-generation and embedded-model options offered GPU and NPU paths.
• Benchmarking: Geekbench AI exposed only its ONNX CPU testing path. It could not provide an accelerated AI benchmark on the reviewed system, so its CPU result did not measure what the Blackwell graphics processor could deliver.
• Model access: Phi Silica, Microsoft’s on-device language model, was not accessible out of the box. Calling its API required developer registration and an associated app identity—extra steps for someone expecting to start experimenting immediately.

The preview also showed stability problems. Indge reported two occasions when browsing AI Dev Gallery’s model selection caused the program to freeze and crash back to the desktop.

Memory changes which models fit

Tom’s Hardware’s same-day review adds a practical constraint to the capacity story. Its 32GB machine arrived with 12GB assigned to the graphics processor, an allocation the reviewer said users could adjust. Memory was tight for AI testing, leading the reviewer to choose smaller models, including Google’s Gemma 4 12B and Qwen 3.6 35B Q2, for use in a tool such as Hermes.

Those are results from different configurations and model choices, not a controlled speed comparison. Ars’s large-model run shows what its higher-memory system accommodated; Tom’s Hardware’s experience shows why buyers cannot treat every Surface Laptop Ultra configuration as offering the same local-AI capacity.

The premium also buys a conventional laptop

Outside AI, Indge praised the display and trackpad. The 15-inch mini-LED touchscreen has a 3270×2180 resolution and refreshes at up to 120Hz. He found it bright and vivid, though its glossy surface produced glare in bright rooms. Those are tangible everyday benefits, separate from whether an AI app supports the graphics hardware.

Gaming also supplied useful evidence of both capability and compatibility limits. Tom’s Hardware ran God of War: Ragnarok at 41–47 frames per second during one fight at 2560×1440, using ultra settings and Nvidia’s DLSS rendering technology. But Red Dead Redemption 2 would not run in its testing. The stronger graphics hardware did not make game support universal.