NVIDIA RTX Spark laptops arrive with up to 128GB for local AI

NVIDIA announced the start of RTX Spark laptop availability on October 16, 2026, with a rollout across ASUS, Dell, HP, Lenovo, Microsoft and MSI. The new Windows platform puts local AI alongside graphics and creative workloads, rather than treating the laptop primarily as a window into a hosted assistant. NVIDIA confirmed the launch in its official announcement .

Up to 128GB of shared memory

RTX Spark combines a Grace CPU and Blackwell RTX GPU with up to 128GB of unified memory. The architecture gives compatible models access to a larger shared memory pool, making it possible to keep more AI computation on the machine.

The maximum specification is not a promise for every configuration. In its Surface launch details , Microsoft notes that memory is dynamically shared between the CPU and GPU, and the amount available to the GPU is less than the total. The advertised one-petaflop AI figure is theoretical FP4 performance using sparsity, not a measured speed for every model.

A 74B model designed for local agents

The hardware launch follows NVIDIA's October 15 release of Nemotron 3.5 Bolt 74B-A9B 24GB . The model uses a mixture-of-experts design with 74 billion total parameters and nine billion active per token. NVIDIA says its native 2-bit quantization is designed to fit the model into approximately 24GB of memory for local, multistep agent workloads.

There is also a practical software entry point: LM Studio lists Bolt with text and image input, reasoning controls, tool use and structured output. That gives developers a way to experiment with a downloadable model rather than starting with a proprietary cloud API. A model's nominal memory footprint should not be confused with the total memory needed for Windows, an inference engine and other applications.

Local does not mean every task stays offline

NVIDIA's software rollout includes easier local-model setup in compatible agents and hybrid routing between on-device and hosted models. The distinction matters: running inference locally can reduce cloud dependence, but web research, connected services and a deliberate fallback to a hosted model still involve external systems.

The rollout also includes Microsoft Execution Containers and NVIDIA OpenShell integrations to control agent access. Availability differs by application: NVIDIA says OpenClaw and GitHub Copilot already support Execution Containers, while the expanded Hermes integration is scheduled for later in October. Local deployment does not remove the need to review permissions.

Premium hardware, not a free replacement for cloud AI

Microsoft announced a starting U.S. MSRP of $2,599 for Surface Laptop Ultra, with availability beginning October 16. That is an entry price, not the price of the maximum-memory model. Other manufacturers have their own configurations and rollout schedules.

The launch does not establish that these laptops match the strongest hosted models on every task. Its significance is the combination of hardware, downloadable models and agent software in a portable Windows system. The practical test is whether a particular local setup completes useful work at an acceptable speed and cost, not simply how many parameters fit into memory.

Source: NVIDIA