Reviewers find Nvidia's RTX Spark well ahead on local AI and behind the top chips on CPU

The first job the Tom's Guide reviewer gave the machine was The Sims 4. The reviewer installed the Hermes agent with the Qwen 3.8 Flash Next model and told it to build a house and get a job ^[10]. About three hours later it had a home and had started a kitchen fire ^[10]. In between it repeated the same mistake, forgetting that the coordinate graph moves with the character, and went round in circles ^[11].

Nvidia's pitch, as Windows Central relays it, is that RTX Spark is the first Windows chipset designed from the ground up to run local AI locally and securely ^[18]. What the reviewer did on day two was smaller. After wiring the agent to a Google Workspace API and a browser script, a single morning command, "check-emails", produces a prioritized to-do list from two inboxes and Slack ^[12]. In our view that is the job a team can evaluate, because it repeats daily and can be checked against a result the user already knows.

Hermes itself runs on Macs and on other Windows PCs, including AMD machines ^[13]. Tom's Guide puts the Spark's edge down to CUDA, and says an overwhelming majority of open-source models used for local AI are built, trained and initially optimized for it ^[14]. The Tom's Guide piece does not give a figure for the speed gap. It sends readers to a live blog for benchmarks ^[4].

On the CPU, Windows Central describes a competitive part. It puts the N1X level with the Apple M4 Pro and Snapdragon X2 Elite, behind the M4 Max and X2 Elite Extreme, and says it does not lead its class ^[5]. In Cinebench 2026 the 20-core part scored 558 single-core and 6,885 multi-core, about 7.5% above the first-generation Snapdragon X Elite's single-core 519 [c13, d2]. Apple keeps the lead in peak single-thread work ^[6].

Qualcomm is already in that comparison with two generations of Snapdragon, so what Spark adds to Windows is Nvidia as a partner. Microsoft worked closely with Nvidia, and Windows Central says vendors including Adobe and Blender have optimized for the chip ^[17].

Windows Central's reviewer wrote: "If you're not a mobile content creator, software developer, or utilize local LLMs often, RTX Spark probably isn't for you." ^[16] Outside AI, the same piece says the part matches or approaches dedicated RTX 4070 and 5060 class GPUs in most workloads ^[7]. The smaller GPU option has 5,120 cores, 16.7% fewer than 6,144 ^[20], and the Windows Central numbers come from the top configuration ^[1].

We'd pilot Spark for the people the six machines target, high-end creators and developers ^[8], and only where local models run most days. The tradeoff is that the pilot buys ahead of battery and thermal results ^[1]. To sort the rest of a refresh, ask two things of each group: do they run local models most days, and is their slowest step waiting on the GPU or memory, or on single-thread CPU speed.

- Local models, GPU or memory bound: the one group where both reviews report a gain [c10, c16]. Pilot here. - Local models, single-thread bound: Apple keeps the peak single-thread lead ^[6], so run the same agent on a MacBook Pro too. - No local models, GPU bound: Spark is level with RTX 4070 and 5060 class GPUs outside AI ^[7], so the chassis and the price decide. - No local models, CPU bound: neither review gives a reason to move.