NVIDIA Introduces DGX Spark 64GB Desktop AI Supercomputer Starting at $4,999

NVIDIA has introduced a new 64GB configuration of its DGX Spark personal AI supercomputer, expanding options for developers, researchers, and creators looking to build and deploy local AI workloads without depending on cloud infrastructure. Scheduled to launch on Friday, October 23, the system starts at $4,999 and will be available exclusively through hardware partners Acer, ASUS, Dell, Gigabyte, HP, and MSI.

The new edition retains the architecture of the 128GB model, including the GB10 Grace Blackwell Superchip, unified memory, DGX OS, and the complete NVIDIA AI software stack.

On-Device Inference for Up to 100-Billion-Parameter Models

The DGX Spark 64GB system is built to run autonomous AI agents, fine-tuning tasks, data science pipelines, and edge deployments natively. A single 64GB unit can run open models with up to 100 billion parameters directly on the device.

Out of the box, the system integrates the NVIDIA Agent Toolkit, CUDA-X AI libraries, and Nemotron open models. It supports key developer inference frameworks and runtimes, including llama.cpp, Ollama, vLLM, LM Studio, and PyTorch with CUDA. On the creative software side, 3D creation suite Blender is among the first major applications adding platform support via an upcoming downloadable installer.

Dual-Unit Clustering via NVIDIA Sync

For workflows that outgrow a single machine, DGX Spark includes an integrated NVIDIA ConnectX-7 network interface card (NIC). Developers can directly connect two systems using a QSFP cable over a 200 GbE fabric without needing dedicated server infrastructure.

Clustering two 64GB units pools their memory to 128GB, doubling memory bandwidth and enabling on-device execution of models with up to 200 billion parameters. In NVIDIA's testing with Qwen 3.8 27B, a clustered two-node setup achieved up to a 1.7x performance improvement over a single unit.

The clustering process is managed by the NVIDIA Sync app and its Cluster Assistant feature, which automatically detects hardware connections, validates system settings, and configures network routing. Because both nodes share an identical software stack, existing single-unit environments scale to two units without software reconfiguration.

Software Ecosystem and Availability

NVIDIA plans to release the NVIDIA Sync Model Launcher at the end of October. The tool will enable one-click downloading and deployment of Qwen3.8 27B across single or multi-node setups, exposing the model to local client laptops and setting up the OpenCode development tool directly in web browsers.

NVIDIA also noted that manufacturer partners—including Acer, ASUS, Dell, HP, Lenovo, Microsoft, and MSI—are rolling out new Windows PCs powered by NVIDIA RTX Spark this month. Meanwhile, Alibaba's lightweight open-weight Qwen-Image-2.1 model is now supported locally across NVIDIA RTX GPUs, DGX Spark, and DGX Station platforms for image generation and editing tasks.

The DGX Spark 64GB configuration will be available starting October 23 through authorized manufacturing partners.