NVIDIA RTX Spark N1X Runs a Local Qwen3.8 27B Agent in 18.7 Minutes

A laptop built on NVIDIA RTX Spark N1X has completed its first AA-AgentPerf-Local test from Artificial Analysis. The prototype 15-inch Surface Laptop Ultra, with 128 GB of unified memory and a price of $5,899.99, ran the local agent workload on Qwen3.8 27B in 18.7 minutes. That is 2.0 times faster than a MacBook Pro with M5 Pro (37.6 minutes) and 1.8 times faster than AMD Ryzen AI Halo (34.5). The laptop trails DGX Spark (17.3 minutes) by only 8%, even though the chip runs Windows on Arm and fits within 45–80 W.

What exactly was measured

AA-AgentPerf-Local is an open tool that replays recorded agent sessions against a local model server. Each request carries the full conversation history, and the model generates exactly as many tokens as the original recording contains. Every system does the same work. Tool execution time is not part of the measurement; only the hardware's response on each turn counts.

The standard workload looks like this:

• 8 tasks: five from PinchBench and three from SWE-bench;
• 168 model turns and 172 tool calls;
• context of up to 56K tokens at its longest point;
• 2.6 million prompt tokens, of which only 196K are new and the rest are repeats;
• 31K output tokens, identical on all systems.

With this much repetition, caching decides the outcome. A server with a cache only needs to read the starting prompt and each turn's increment: a median of about 450 tokens, and a maximum of 6.4K after a large tool output.

Results by model

N1X is effectively the same GB10 that sits in DGX Spark: 20 Arm cores (10 Cortex-X925 and 10 Cortex-A725), a Blackwell GPU with 6,144 CUDA cores, and up to 128 GB of LPDDR5X. The laptop's memory is faster (9400 vs. 8533 MT/s), with bandwidth of 300 GB/s vs. 273. But its thermal envelope is lower: 45–80 W vs. 140 W for the GB10. That was enough to nearly catch the desktop box.

• Qwen3.8 27B (dense, 27B active): laptop 18.7 min, DGX Spark 17.3, Ryzen AI Halo 34.5, MacBook Pro 37.6, RTX 5090 4.9.
• Qwen3.6 35B A3B (3B active): 7.6 min vs. 7.3 for DGX Spark, 11.7 for Ryzen AI Halo, and 11.9 for MacBook Pro. The RTX 5090 finishes in 2.0 minutes.
• Ling 3.0 Flash (124B, 5.1B active): 15.3 min vs. 14.9 for DGX Spark and 25.1 for Ryzen AI Halo.

The last line shows the main advantage of 128 GB. Ling 3.0 Flash's weights take up 77 GB, so they fit neither in the MacBook Pro's 64 GB nor in the RTX 5090's 32 GB of VRAM. Wherever a model fits in 32 GB, the desktop 5090 is fastest: on Qwen3.8 27B it beats the laptop by roughly 3.8 times.

Recipes, prices, and caveats

Each system ran with its own recipe. The laptop and DGX Spark use llama.cpp with UD-Q4_K_M weights and DFlash speculative decoding with 7 draft tokens. The MacBook Pro, Ryzen AI Halo, and RTX 5090 use MTP (4 tokens on the Mac, 2 on the others). The operating systems differ as well: Windows 11 on Arm on N1X, Linux on DGX Spark and AMD, macOS on the Mac. The comparison shows ready-made stacks, not raw hardware. Artificial Analysis itself warns that it is testing other weight formats (including IQ4_XS) and Splash, a Metal engine for Apple Silicon with DFlash. Early measurements on some systems are faster, so the picture may shift.

The prices in the test are approximate: about $3,700 for the MacBook Pro, $4,000 for Ryzen AI Halo, about $6,000 for DGX Spark, and about $9,500 for an RTX 5090 build (a $7,000 card plus $2,500 for the rest of the PC). At $5,900, the laptop costs more than the Mac and the AMD system but less than DGX Spark.

Who this is for

If an agent needs to run locally and on the road, the N1X laptop is the only option with 128 GB of memory and a battery. It handles 124B-class MoE models that a 64 GB Mac simply cannot load, and it delivers DGX Spark speed. If the machine sits on a desk and the models fit in 32 GB, the RTX 5090 is the better bet: it is many times faster, though it costs the most of all.

For models under 64 GB and a budget of around $3,700, the MacBook Pro remains, but agent tasks take twice as long on it. Before buying, recalculate the prices for your region: the benchmark's leaderboard lets you enter your own. Keep in mind that the test ran on a prototype in high-performance mode on mains power, and results on battery may differ.