NVIDIA's RTX Spark laptops gain new models, local AI models and agent GPU controls

NVIDIA used an October 28 blog post to widen its RTX Spark line: Windows laptops built around the chip are now shipping from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599, with 128GB unified-memory configurations already on shelves. The announcement paired the hardware with new local AI models, GPU controls for background agents and a batch of game updates.

On the hardware side, RTX Spark puts a Blackwell RTX GPU and a Grace CPU on a single superchip with up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance. NVIDIA says the machines are as thin as 14mm and as light as three pounds, and because CPU, GPU and memory draw from one pool, a laptop can take on creative and model workloads that previously called for a desktop workstation.

The most substantive model update is Nemotron 3.5 Transcribe, an open-weight speech model that handles transcription, speaker attribution and translation from a single prompt across 40 languages, seven of them Indic. NVIDIA says it runs on roughly 4.3 billion active parameters, 80% fewer than the top open-weight model, ranks first on Hugging Face's Open ASR leaderboard and second among open-weight models on Artificial Analysis. Weights are on Hugging Face in BF16, FP8 and NVFP4 formats, with support for RTX Spark, DGX Spark, Jetson AGX Thor and GeForce RTX 5090.

The agent-related change lands in the NVIDIA App. A new resource management control lets users decide how the GPU is shared: prioritize the foreground application for smooth performance, or give more compute to an agent running in the background so it finishes sooner. The feature lives under System > Advanced and requires hardware-accelerated GPU scheduling to be enabled in Windows.

NVIDIA also gave Liquid AI's LFM3 family day-zero support. The line spans text, vision, audio, omni and decision models on one shared backbone, released as open weights, and NVIDIA claims the new 1.7-billion-parameter base model outperforms its previous 2.6-billion-parameter base. Developers can run the models on RTX Spark through llama.cpp and vLLM, and fine-tune them locally.

On the gaming side, S-GAME's Phantom Blade Zero launches October 28 at 7 p.m. PT on Steam and the Epic Games Store with ray-traced effects added in collaboration with NVIDIA, plus DLSS 4.5 dynamic multi frame generation at up to 6x. RTX Remix 1.6, released the previous week, brings path tracing and DLSS to Minecraft Java Edition 1.21.1 and adds native RTX Spark support along with agent-assisted modding workflows.

Taken together, the pitch is no longer just a laptop that can run models locally, but a personal workstation where agents can live for long stretches. When a background agent and a foreground app compete for the same GPU, allocating compute stops being a technical detail and becomes a product decision, which is why NVIDIA is now putting that switch in front of users.

Three things are worth watching: real street prices and supply for the 128GB configurations, battery life and thermals once agents run persistently, and whether local models like Nemotron 3.5 Transcribe and LFM3 hold their leaderboard accuracy in actual meetings and customer-service work.