NVIDIA RTX Spark N1X 笔记本首份 AA-AgentPerf-Local 测试结果

Artificial Analysis

All articlesOctober 15, 2026First AA-AgentPerf-Local results for an NVIDIA RTX Spark N1X laptop

NVIDIA’s RTX Spark N1X puts up to 128 GB of unified memory and a Blackwell GPU in a Windows laptop. In our first AA-AgentPerf-Local results, an RTX Spark laptop completes our local agent workload on Qwen3.8 27B in 18.7 minutes.Laptops & workstations leaderboardAA-AgentPerf-Local on GitHubThis article shows AA-AgentPerf-Local results as of October 15, 2026. Its charts update live, so they may differ from the figures in the text. For the latest results, see the laptops & workstations leaderboard.

AA-AgentPerf-Local replays recorded agent sessions on local hardware

AA-AgentPerf-Local is our open-source tool for measuring how fast a laptop or workstation serves an AI agent. It replays recorded agent sessions against a model server running on the machine, with each request carrying the whole conversation so far, and measures how long the hardware takes to answer every turn.

The model generates exactly the number of tokens recorded for each turn, so every system does the same work. Most of each prompt is a repeat of the previous prompt’s contents. A server that keeps its cache only has to read the opening prompt of each task and what each turn adds (a median of about 450 tokens, and up to 6.4K after a large tool output).

8

Recorded agent tasks, five from PinchBench and three from SWE-bench

168

Model turns, each carrying the full conversation so far

56K

Tokens of context at the longest point in a task

2.6M

Prompt tokens sent to the model across the workload

196K

New prompt tokens (the rest repeat an earlier prompt)

31K

Output tokens, the same count on every system

AA-AgentPerf-Local default workload

Our leaderboard covers three models at 4-bit precision, from a 27B dense model to a 124B mixture-of-experts model, on the five systems below. The RTX Spark laptop joins NVIDIA DGX Spark, the AMD Ryzen AI Halo, a MacBook Pro (M5 Pro) and a desktop with a GeForce RTX 5090.

Systems

MacBook Pro

M5 Pro · 64 GB · 20-core GPULaunch MSRP (64 GB)$3,000Memory64 GB unifiedMemory bandwidth307 GB/sMemory typeLPDDR5X-9600ChipM5 ProLaunchedMar 2026

NVIDIA DGX SparkLaunch MSRP$4,000Memory128 GB unifiedMemory bandwidth273 GB/sMemory typeLPDDR5X-8533ChipGB10LaunchedOct 2025

NVIDIA RTX Spark N1X

20-core CPU · 6144-core GPULaunch MSRP (128 GB)$5,900Memory128 GB unifiedMemory bandwidth300 GB/sMemory typeLPDDR5X-9400ChipN1XLaunchedOct 2026

Qwen3.8 27B (Xhigh)Intelligence Index34Total parameters27BActive parameters27BContext window256KReleasedAug 2026Builds testedQ4_K_M, UD-Q4_K_M

Qwen3.6 35B A3B (Reasoning)Intelligence Index18Total parameters36BActive parameters3BContext window256KReleasedApr 2026Builds testedUD-Q4_K_M, NVFP4

Ling 3.0 FlashIntelligence Index20Total parameters124BActive parameters5.1BContext window256KReleasedAug 2026Builds testedQ4_K_M

RTX Spark puts DGX Spark's chip in a Windows laptop

RTX Spark is NVIDIA’s platform for Windows on Arm PCs, which NVIDIA pitches as built to run AI agents locally. Its N1X chip is essentially the GB10 from DGX Spark (NVIDIA’s own recipe catalogue lists RTX Spark as GB10), with a 20-core Arm CPU, a 6,144-core Blackwell GPU and up to 128 GB of LPDDR5X memory shared between them. In the laptop the memory runs faster, giving 10% more bandwidth than DGX Spark (300 GB/s against 273 GB/s). However, the chip runs at 45 to 80 W (NVIDIA’s stated range for N1X laptops), where GB10 is rated at 140 W.

We tested a prototype 15-inch Surface Laptop Ultra with 128 GB of memory, priced at $5,899.99. It runs Windows 11 on Arm and serves every model with a native Windows build of llama.cpp on the GPU. We ran it plugged in with Windows set to its high performance power mode.

GPUBlackwell, 6,144 CUDA coresBlackwell, 6,144 CUDA cores20 cores

Memory bandwidth300 GB/s273 GB/s307 GB/s

Chip power45 to 80 W140 WNot published

Operating systemWindows 11 on ArmDGX OS (Ubuntu)macOS

Specifications from NVIDIA’s RTX Spark page and porting guide, NVIDIA’s DGX Spark documentation and Apple as of October 2026, with memory speeds as reported by the systems we tested

Each system runs its own recipe

We run each model with a recipe chosen for each system, using the vendor’s published recipe where one exists, so the runtime, speculative decoding and 4-bit weight format differ between systems. The RTX Spark laptop runs NVIDIA’s llama.cpp recipes for RTX Spark, ported into AA-AgentPerf-Local with NVIDIA’s settings.

Qwen3.8 27BRuntimeWeightsSpeculative decoding

RTX Spark laptopllama.cpp, WindowsUD-Q4_K_MDFlash, 7 draft tokens

DGX Sparkllama.cpp, LinuxUD-Q4_K_MDFlash, 7 draft tokens

MacBook Pro (M5 Pro)llama.cpp, macOSQ4_K_MMTP, 4 draft tokens

AMD Ryzen AI Halollama.cpp, LinuxQ4_K_MMTP, 2 draft tokens

GeForce RTX 5090 desktopllama.cpp, LinuxQ4_K_MMTP, 2 draft tokens

The recipe behind each system’s fastest Qwen3.8 27B result

We are also testing more weight formats and runtimes, 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. We will add their results once they pass our accuracy checks, which we will publish soon. Early tests show faster completion times on some systems.

The RTX Spark laptop finishes the workload in 8-19 minutes, depending on the model

➤On Qwen3.8 27B the RTX Spark laptop completes the workload in 18.7 minutes, close to DGX Spark (17.3), 2.0x faster than the MacBook Pro (37.6) and 1.8x faster than the Ryzen AI Halo (34.5). The GeForce RTX 5090 desktop is fastest, at 4.9 minutes.

➤On Qwen3.6 35B A3B, which activates 3B parameters per token, it takes 7.6 minutes, close to DGX Spark (7.3), 1.5x faster than the Ryzen AI Halo (11.7) and 1.6x faster than the MacBook Pro (11.9).

➤On Ling 3.0 Flash, the largest model on our leaderboard, it takes 15.3 minutes, close to DGX Spark (14.9) and 1.6x faster than the Ryzen AI Halo (25.1). Its 77 GB of weights do not fit on the MacBook Pro or the RTX 5090.

Completion Time by Model and SystemTime to complete the AgentPerf-Local default workload (168 turns, 172 tool calls) · Lower is betterMacBook Pro (M5 Pro, 64 GB)64 GB unifiedAMD Ryzen AI Halo128 GB unifiedNVIDIA DGX Spark128 GB unifiedNewNVIDIA RTX Spark Laptop128 GB unifiedNVIDIA GeForce RTX 509032 GB VRAMQwen3.8 27B (Xhigh)27B parameters37.6 min34.5 min17.3 min18.7 min4.9 minQwen3.6 35B A3B (Reasoning)36B, 3B active parameters11.9 min11.7 min7.3 min7.6 min2.0 minLing 3.0 Flash124B, 5.1B active parametersDoes not fitin 64 GB unified25.1 min14.9 min15.3 minDoes not fitin 32 GB VRAMFasterSlowerDoes not fit in memory

Total time the system takes to answer every turn of the workload's recorded agent tasks. Turns are sent one at a time, each with the conversation so far, and the model generates the same number of tokens as in the original recording, so every system does the same work. Time spent running tools is not included.

Qwen3.6 35B A3B, which activates 3B parameters per token, is the fastest model on every system, and the RTX Spark laptop finishes it in 7.6 minutes against 15.3 for Ling 3.0 Flash and 18.7 for the dense Qwen3.8 27B. The GeForce RTX 5090 desktop is the fastest system on every model that fits in its 32 GB.

Completion Time|Qwen3.8 27B (Xhigh)Time to complete the AgentPerf-Local default workload (168 turns, 172 tool calls) · Lower is better

Total time the system takes to answer every turn of the workload's recorded agent tasks. Turns are sent one at a time, each with the conversation so far, and the model generates the same number of tokens as in the original recording, so every system does the same work. Time spent running tools is not included.

Prices for these systems are approximate and shift from week to week. At current prices, the $5,899.99 RTX Spark laptop costs more than the MacBook Pro (~$3,700) and the Ryzen AI Halo ($4,000) but slightly less than DGX Spark (~$6,000), and the RTX 5090 desktop is the most expensive system at ~$9,500 (a ~$7,000 card plus a conservative $2,500 for the rest of the PC). The chart below uses these prices, and our leaderboard lets you enter your own.

Completion Time vs. Price|Qwen3.8 27B (Xhigh)Time to complete the AgentPerf-Local default workload against the price of the system as tested · Lower time and lower price is betterMost attractive quadrantPareto lineApproximate current prices, October 2026

Total time the system takes to answer every turn of the workload's recorded agent tasks. Turns are sent one at a time, each with the conversation so far, and the model generates the same number of tokens as in the original recording, so every system does the same work. Time spent running tools is not included.

Approximate current price in US dollars, as of October 2026. Systems that still sell at launch MSRP are priced at it. A graphics card is priced at the card plus $2,500 (arbitrary estimate) for the rest of the system. Enter your own prices to reflect the prices available to you.

Explore further

Compare every system and model on the laptops & workstations leaderboard

See the full launch command behind each result on the configurations page

Run any recipe on your own machine with AA-AgentPerf-Local on GitHub$/$

NVIDIA Vera Rubin NVL72 serves up to ~5.7x the total throughput per megawatt of GB300 NVL72 on AA-AgentPerf, with first results on DeepSeek V4 Pro, DeepSeek V4.1 Flash, GLM-5.2 and Kimi K3.

October 15, 2026

NVIDIA Releases Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B A9B

Nvidia’s Nemotron 3.5 Super VL is now the leading US open weight model, scoring 26 on the Artificial Analysis Intelligence Index

October 15, 2026

Welcome to the era of embodied AI: Introducing AA-Robotics v0.1

We are launching AA-Robotics v0.1, our research preview that measures frontier language model performance on understanding, planning, and acting in the physical world. In v0.1, each model is given control of a 6-axis robot arm (5 joints and a gripper) and a simple objective. We’re releasing results for our first eight tasks today.

October 14, 2026