RTX Spark laptops from $2,599 now let you split the GPU with background agents

Nvidia says its RTX Spark laptops are on sale from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599. In an official post, the company also announced a control for sharing the GPU between apps and agents, plus new models and games.

One chip, with memory scaling to 128GB

According to Nvidia, each RTX Spark laptop puts a Blackwell RTX GPU and a Grace CPU on a single superchip. The memory is unified and tops out at 128GB, and the company quotes a ceiling of one petaflop of FP4 AI compute. Because creative projects and AI models draw on one memory pool, the company says a laptop can handle jobs that used to need a desktop workstation, even on battery.

Nvidia describes bodies as slim as 14mm and weighing about three pounds, and it promises a full day of battery. Options include touchscreens, pen support and two-in-one convertibles. The 128GB configurations are already on shelves, the company says. Adobe Photoshop and Premiere have native RTX Spark support in beta releases.

A dial for foreground apps against background agents

The most practical addition is a resource management feature in the NVIDIA App. Users can favor the app they are working in, or give a long-running background agent more GPU compute so it finishes sooner. The setting sits under System > Advanced and requires hardware-accelerated GPU scheduling to be switched on in Windows Settings.

Nvidia gave no figures on how much each setting shifts performance, so the effect on a game or render while an agent runs is not established by the announcement.

New models and games arrive alongside

Nemotron 3.5 Transcribe, an open weight speech model, is on Hugging Face today. Nvidia says one promptable model handles transcription, speaker labels, translation and questions about the audio across 40 languages. The company says it holds first place on the Hugging Face Open ASR board and second among open weight models at Artificial Analysis, using about 4.3 billion active parameters. Those rankings are Nvidia’s own claims. Quantized versions that run locally on RTX Spark are listed as coming soon.

Liquid AI’s LFM3 family has day-zero support. It spans text, vision, audio, omni and decision models, and Nvidia says it runs natively through llama.cpp and vLLM, with local fine-tuning possible. Liquid AI says its new 1.7-billion-parameter base model beats its earlier 2.6-billion-parameter one.

On the gaming side, S-GAME’s Phantom Blade Zero went on sale the evening of Oct. 28 (7 p.m. PT) through Steam and the Epic Games Store, with ray tracing and DLSS 4.5. Nvidia says DLSS Dynamic Multi Frame Generation can reach up to 6x frame generation. RTX Remix 1.6 brings path tracing to Minecraft Java Edition 1.21.1, and Portal with RTX and the Half-Life 2 RTX demo now support RTX Spark.

A price that sits below Apple’s new MacBook Pro

The $2,599 starting price puts RTX Spark in the premium tier, and it lands next to Apple’s latest pro laptop. Our earlier coverage reported that Apple’s new MacBook Pro starts at $2,899. That is $300 more than Nvidia’s entry point, though the two lists do not say which configurations are comparable. Nvidia has not said what the $2,599 model includes, and the 128GB machines will cost more.

Nvidia’s argument is different from Apple’s: CUDA. The company says developers can run models and agents locally on the same platform that powers its cloud AI, and deploy to more than 150 million RTX AI PCs. That claim has not been independently verified, and it matters only to people whose code already depends on CUDA.

The models also fit the local pitch. Our earlier report on Nemotron 3.5 Bolt E4B described a 4GB model for local assistants, game characters and robots. Nvidia’s post lists it as available on RTX Spark, and a larger Bolt 74B-A9B 24GB model is aimed at heavier local agents. For the speech model, our coverage of Nemotron 3.5 Transcribe noted that tuning cut medical errors by 62%, a result that supports the specialized-vocabulary use Nvidia describes.

What is still missing

The announcement does not list prices per model or per maker, nor battery life figures. It gives no benchmarks for running a large model unplugged. Nvidia also points to a separate Microduck robot demo and a 64GB DGX Spark now sold by partners, but neither changes what a laptop buyer gets today.