NVIDIA RTX Spark N1X Laptop Completes Local Agent Workload in 18.7 Minutes

NVIDIA RTX Spark N1X Laptop Completes Local Agent Workload in 18.7 Minutes

NVIDIA’s RTX Spark N1X brings a Blackwell GPU and up to 128 GB of unified memory to a Windows laptop. In the first AA-AgentPerf-Local results for the platform, an RTX Spark laptop completed the local agent workload on Qwen3.8 27B in 18.7 minutes.

The results described here reflect testing as of October 15, 2026. The benchmark charts update live, so current figures may differ.

AA-AgentPerf-Local replays recorded agent sessions

AA-AgentPerf-Local is an open-source tool that measures how quickly a laptop or workstation serves an AI agent. It replays recorded agent sessions against a model server running locally. Each request includes the complete conversation so far, and the benchmark records the time required to answer every turn.

The model generates exactly the number of tokens recorded for each turn, ensuring that every system performs the same work. Most prompt tokens repeat content from earlier turns. A server that maintains its cache needs to process only the task’s initial prompt and the new material added at each turn. The median addition is approximately 450 tokens, rising to as many as 6.4K tokens after a large tool output.

The default workload contains:

• Five recorded agent tasks from PinchBench and three from SWE-bench
• 168 model turns and 172 tool calls
• The full conversation history sent with each turn
• Context reaching up to 256K tokens, depending on the task
• The same output-token count for every system

Tested systems and models

The leaderboard covers three models using 4-bit precision, ranging from the 27B dense Qwen3.8 27B to the 124B mixture-of-experts Ling 3.0 Flash. Testing includes five systems:

The listed launch MSRPs are approximately $3,000 for the 64 GB MacBook Pro, $4,000 for the AMD Ryzen AI Halo, $4,000 for NVIDIA DGX Spark, $5,900 for the 128 GB RTX Spark N1X laptop, and $2,000 for the RTX 5090 card.

The models tested are:

RTX Spark brings the DGX Spark chip to a Windows laptop

RTX Spark is NVIDIA’s platform for Windows on Arm PCs. Its N1X chip is essentially the GB10 used in DGX Spark. NVIDIA’s recipe catalogue identifies RTX Spark as GB10. The chip includes a 20-core Arm CPU, a 6,144-core Blackwell GPU, and up to 128 GB of LPDDR5X memory shared between the CPU and GPU.

The laptop configuration uses faster memory than DGX Spark, providing 300 GB/s of bandwidth compared with 273 GB/s. However, the N1X operates at 45 to 80 W, NVIDIA’s stated range for N1X laptops, while the GB10 is rated at 140 W.

Testing used a prototype 15-inch Surface Laptop Ultra with 128 GB of memory and a price of $5,899.99. The system ran Windows 11 on Arm and served every model with a native Windows build of llama.cpp on the GPU. Testing was performed while plugged in, with Windows set to its high-performance power mode.

Each system uses a separate software recipe

Each model runs with a recipe selected for the relevant system. Where available, the recipe comes from the vendor. As a result, runtimes, speculative-decoding settings, and 4-bit weight formats vary.

The RTX Spark laptop uses NVIDIA’s llama.cpp recipes for RTX Spark, ported into AA-AgentPerf-Local with NVIDIA’s settings.

Additional weight formats and runtimes are also being tested, including smaller 4-bit formats such as IQ4_XS and Inco AI’s Splash, a Metal engine for Apple Silicon that uses DFlash speculative decoding. Those results will be added after they pass accuracy checks.

RTX Spark completes the workload in 7.6 to 18.7 minutes

The benchmark results for the default workload are:

On Qwen3.8 27B, the RTX Spark laptop completes the workload in 18.7 minutes. That is close to DGX Spark’s 17.3 minutes, and faster than the MacBook Pro at 37.6 minutes and the Ryzen AI Halo at 34.5 minutes. The GeForce RTX 5090 desktop is fastest at 4.9 minutes.

On Qwen3.6 35B A3B, which activates 3B parameters per token, the RTX Spark laptop takes 7.6 minutes. DGX Spark takes 7.3 minutes, while the Ryzen AI Halo and MacBook Pro take 11.7 and 11.9 minutes, respectively.

On Ling 3.0 Flash, the largest model in the benchmark, the RTX Spark laptop takes 15.3 minutes, compared with 14.9 minutes for DGX Spark and 25.1 minutes for the Ryzen AI Halo. The model’s 77 GB of weights do not fit in the MacBook Pro’s 64 GB of unified memory or the RTX 5090’s 32 GB of VRAM.

Qwen3.6 35B A3B is the fastest model on every system. On the RTX Spark laptop, it completes the workload in 7.6 minutes, compared with 15.3 minutes for Ling 3.0 Flash and 18.7 minutes for the dense Qwen3.8 27B. The GeForce RTX 5090 desktop is the fastest system for every model that fits within its 32 GB of memory.

Completion time and system price

Completion time is the total time required to answer every turn in the recorded agent tasks. Turns are sent one at a time with the conversation history included, and each model generates the same number of tokens on every system. Time spent running tools is excluded.

Prices vary over time. Based on the approximate prices used for the October 2026 comparison, the $5,899.99 RTX Spark laptop costs more than the MacBook Pro at approximately $3,700 and the AMD Ryzen AI Halo at $4,000. It costs slightly less than DGX Spark at approximately $6,000. The GeForce RTX 5090 desktop is estimated at approximately $9,500, based on a roughly $7,000 graphics card and an estimated $2,500 for the rest of the system.

The price comparison uses these estimates. A separate comparison can use different prices to reflect local availability or current system costs.

Benchmark scope

AA-AgentPerf-Local measures the time required for the systems to generate the recorded responses. It does not measure tool-execution time, and it does not imply that the systems perform identical work outside the recorded workload. Results can change as runtimes, model builds, drivers, and hardware configurations change.