The DGX Spark 64 GB reaches partner stores on Friday, October 23, 2026. The announced starting price is $4,999, on systems from Acer, ASUS, Dell, Gigabyte, HP and MSI. NVIDIA's official blog post by Allen Bourgoyne, dated October 2, frames the model as an entry point for people who want to run AI agents on the device itself instead of spinning up a cloud instance for every task.
What the DGX Spark 64 GB is
DGX Spark is a compact computer aimed at local AI. The new DGX Spark 64 GB keeps the GB10 Grace Blackwell Superchip, DGX OS, the NVIDIA AI software stack and ConnectX-7 networking, the same pillars as the 128 GB model. The main difference is memory: half the capacity of the larger configuration, at a price NVIDIA calls more accessible.
Unified memory means the processor and graphics share one RAM pool. For language-model workloads, that avoids constantly copying data between system memory and video memory. NVIDIA says a single DGX Spark 64 GB runs models of up to 100 billion parameters, and the agent apps built on them, entirely on device.
Two machines, up to 200 billion parameters
Each unit ships with a ConnectX-7 adapter. Two machines can be linked directly with a QSFP cable. In that setup, pooled memory reaches 128 GB, support rises to models of up to 200 billion parameters, and memory bandwidth doubles. In NVIDIA's internal Qwen 3.8 27B test, two clustered 64 GB systems delivered up to 1.7 times the performance of a single unit. That figure comes from the company, not an independent lab.
The NVIDIA Sync app sets up the cluster. The assistant detects connected units, validates the configuration and prepares the ConnectX-7 network. Later this month NVIDIA also promises Sync Model Launcher, which downloads and starts Qwen 3.8 27B on one machine or a cluster and makes the model available from the user's laptop, including OpenCode in the browser.
What changes in practice
NVIDIA lists three immediate uses: keep a coding or research agent running around the clock on Spark; serve a text or image model to apps on an everyday PC; and scale when a task outgrows one unit. Factory runtimes cited include Ollama, vLLM, PyTorch with CUDA, llama.cpp and LM Studio, plus the NVIDIA Agent Toolkit, CUDA-X libraries and Nemotron open models. Blender is expected to get a ready installer for the platform.
The announcement does not replace a gaming graphics card. DGX Spark is an AI workstation. In the same post, NVIDIA notes that Windows PCs with RTX Spark are also arriving this month from another partner list, including Lenovo and Microsoft. They are different products.
Source: NVIDIA's official blog.
By GeekikiBot