The first laptops with Nvidia’s RTX Spark combo chip went on sale on October 16, and with them come the first independent measurements. They paint a more nuanced picture than the launch event did: for AI agents that process large amounts of text, the chip clearly beats a comparable MacBook Pro, but not when it comes to raw text output from large models. If you are considering running language models on your own laptop, the tests offer useful guidance.
Key takeaways
- Artificial Analysis measured roughly half the completion time of a MacBook Pro with M5 Pro and 64 GB on a series of agent tasks, using a Surface Laptop Ultra with 128 GB.
- c’t 3003, the video format of German tech magazine c’t, by contrast found the M5 Pro ahead on text output from a dense 27-billion-parameter model; the Asus ProArt P14 reached about 14.5 tokens per second.
- Mixture-of-experts models run fast: c’t measured almost 61 tokens per second, StorageReview up to 89.
- Memory is fixed at purchase; in Germany, the ProArt P14 ranges from about €3,300 (32 GB) to €6,700 (128 GB).
- Weak spots: a fan that c’t says spins up even at idle, and unexplained power cuts under full load according to StorageReview.
What the tests measure and why they seem to contradict each other
RTX Spark, internally called N1X, combines a Blackwell GPU with an Arm processor, and both share up to 128 GB of unified memory. Technically, it is essentially the chip from Nvidia’s DGX Spark mini computer, but with far lower power draw: Artificial Analysis cites 45 to 80 watts in the laptop versus 140 watts in the desktop original. We covered the announcement in early October, along with Microsoft’s sandboxes for agents. Back then, it was unclear how fast the chip actually is.
On October 15, Artificial Analysis ran its own agent benchmark on a preproduction Surface Laptop Ultra with 128 GB: eight recorded agent tasks, including three from the SWE-bench coding test, with 168 model turns and 172 tool calls. With the dense model Qwen3.8 27B, the laptop took 18.7 minutes, the MacBook Pro with M5 Pro and 64 GB 37.6 minutes, and a system with AMD’s Ryzen AI Halo 34.5 minutes. Nvidia’s DGX Spark came in slightly ahead of the laptop at 17.3 minutes, while a desktop with a GeForce RTX 5090 was far ahead at 4.9 minutes.
The c’t 3003 team reached a different result with the Asus ProArt P14 with 32 GB: a dense Qwen model with 27 billion parameters runs at about 14.5 tokens per second, slower than on the M5 Pro. The contradiction disappears once you look at the type of work. Agents read a lot and write little: in the Artificial Analysis benchmark, about 2.6 million input tokens face only 31,000 generated ones. Reading input, known as prompt processing, is compute-intensive, and c’t rates RTX Spark ahead of Strix Halo and the M5 Pro here. Generating new words, on the other hand, depends mainly on memory bandwidth, where Apple has the edge.
The numbers are not fully comparable. Artificial Analysis stresses that each system ran its own software setup; on the Nvidia laptop, for example, a technique called speculative decoding helped, in which a small helper model proposes words. The memory configurations of the test machines also differ.
What the laptop is good for in daily use
RTX Spark performs best with so-called mixture-of-experts models, which activate only a small share of their parameters per word. Qwen3.6 35B A3B, with 36 billion parameters of which 3 billion are active at a time, reached almost 61 tokens per second at c’t without extra tricks. StorageReview measured 89 tokens per second with the same model in a coding run via the Codex command line, and about 58 with gpt-oss:20b. That is enough for smooth work with a local coding assistant, with no cloud subscription and without code or data leaving the machine.
That said, in the c’t coding test, the slower 27B model built the more sensible website than the speedy 35B model. With local AI, speed is only half the story. And large models need space. With 32 GB, c’t had to switch the 27B model to a more heavily compressed version to run it alongside Photoshop and would have preferred 48 GB. The Ling 3.0 Flash model, with 124 billion parameters, takes up 77 GB and ran in the Artificial Analysis test only on systems with enough memory, not on the 64 GB MacBook and not on the RTX 5090 with 32 GB.
The chip scores on efficiency: c’t measured about 40 watts under sustained load with peaks around 60 watts, while a Strix Halo system drew about 120 watts. During video playback, the ProArt P14’s battery lasted more than 19 hours. On the downside, according to c’t, the fan occasionally spins up even at idle. Games run through Microsoft’s Prism translation layer; Cyberpunk 2077 reached 63 frames per second at 1080p with Ultra settings and no upscaling.
The open issues
StorageReview reports a serious problem: a Surface Laptop Ultra with 128 GB shut down hard three times in one evening under sustained full load on all processor cores, and a previous review unit had failed the same way. According to the testers, the logs point to power delivery or firmware protection, not overheating. Whether an update will fix it is unclear. Linux still takes manual work too: c’t got Ubuntu for Arm running only with a custom kernel and help from Claude Code, but once it did, the MoE model ran 9 to 14 percent faster than under Windows.
Who RTX Spark is worth it for
The most important decision comes at purchase, because the unified memory cannot be expanded later. If you want to run local agents or larger models, lean toward 64 or 128 GB; the 32 GB version is more of a fast work machine with an AI option than an AI machine. In Germany, the ProArt P14 costs about €3,300 with 32 GB, €4,300 with 64 GB, and €6,700 with 128 GB, according to c’t. If you only use models under 24 GB, StorageReview says a laptop with RTX 5090 graphics is much faster. And if a low-cost entry matters most, c’t points to Strix Halo laptops under €2,000 as offering plenty of performance for the money. Our take on the DGX Spark with 64 GB also helps if it does not have to be a laptop.
Conclusion
RTX Spark is no MacBook killer, but it is the first Windows laptop that can seriously handle local AI agents. Its strength lies exactly where agents spend their time: reading long contexts, holding large models in memory, and running Nvidia’s CUDA software, which tools like ComfyUI depend on. Apple stays ahead on raw writing speed and fan control, and the unexplained shutdowns are a reason for buyers to wait until Microsoft and Nvidia have shipped fixes. If you buy today, you are mainly buying memory. The real message of the tests: for local AI, the workload decides, not the spec sheet.