RTX Spark laptops: local AI and agent controls

NVIDIA says RTX Spark laptops are now sold by ASUS, Dell, HP, Lenovo, Microsoft and MSI. Each machine pairs a Blackwell RTX GPU with a Grace CPU on one superchip, with up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance.

NVIDIA describes the setup as putting the CUDA platform, RTX graphics and the capabilities of a personal AI supercomputer inside premium Windows laptops, aimed at developers, creators and gamers.

The company also lists thin designs: as slim as 14mm and as light as three pounds, with all-day battery life claimed. Those are manufacturer figures, not results from independent testing.

NVIDIA introduced Nemotron 3.5 Transcribe, an open weight speech model that handles transcription, speaker labels and translation in one promptable system. It covers 40 languages, including seven Indic languages, and users can ask it plain-language questions about a recording.

NVIDIA says the model ranks first on the Hugging Face Open ASR leaderboard and second among open weight models on Artificial Analysis. It reports roughly 4.3 billion active parameters, about 80% fewer than the top open weight model, which the company says makes local transcription more practical.

The weights are open, so developers choose where audio is processed. Quantized versions for local hardware are described as coming soon, not available yet.

Long-running agents and everyday apps share the same GPU on RTX Spark. A new control in NVIDIA App lets you decide the split: favor the foreground app for smoother performance, or give a background agent more resources so it finishes sooner.

The setting lives under System > Advanced. It requires hardware-accelerated GPU scheduling to be turned on in Windows Settings, so check that first if the option looks unavailable.

NVIDIA describes the balance as depending on what matters to the user at that moment, and does not publish measured performance differences for either setting.

Liquid AI's LFM3 family spans text, vision, audio, omni and decision models built on one shared backbone, with base and post-trained versions released as open weights. NVIDIA says the new 1.7-billion-parameter base model outperforms Liquid AI's previous 2.6-billion-parameter base model.

NVIDIA lists local uses such as triaging inboxes, managing calendars, reading documents and app screens, transcribing speech and routing support tickets. Developers can run LFM3 on RTX Spark through llama.cpp and vLLM, and fine-tune locally.

These are capabilities NVIDIA describes for the models, not results from hands-on testing.

On the developer side, NVIDIA describes workflows where people direct AI agents in natural language to assemble simulation scenes. Omniverse libraries supply GPU-accelerated physics, rendering and sensor simulation, while developers review results and steer changes.

One example: David Wehr, from NVIDIA's simulation technology team, used an AI model to reconstruct a building from panoramic photos and point-cloud measurements. The output is an editable model where shelving and machinery can be moved to test warehouse layouts before physical changes.

Another: Bhavik Sanganee, a principal engineer at NVIDIA, built a warehouse simulation that held 30 frames per second at full HD with 80 conveyors, 40 arms and 800 boxes over a two-minute test. Doubling the scene again cut performance, and testing traced the slowdown to the physics and robot-control pipeline.

NVIDIA's figures on memory and performance come from the company, and the quantized builds needed for local speech work are still listed as coming soon.

If you rely on background agents, the resource control is the feature to look at, and it depends on a Windows setting you may need to enable. For simulation work, the documented examples show agents helping assemble scenes, with performance limits still showing up as scenes grow.

I haven't tested these laptops, so I can't tell you how they feel in daily use. The announcements describe configurations and workflows; they do not provide measured results you can compare across machines.