NVIDIA adds a 64GB DGX Spark for local AI, with two-unit clustering

NVIDIA announced on October 2, 2026 that DGX Spark, its compact desktop system for local AI work, will come in a 64GB unified-memory configuration. The 64GB model is sold only through partners Acer, ASUS, Dell, Gigabyte, HP and MSI, from Friday, Oct. 23, starting at $4,999. NVIDIA paired the launch with NVIDIA Sync Cluster Assistant, a tool for linking two units. A one-click model launcher is due at the end of the month.

What is new in DGX Spark 64GB

The new configuration halves the memory of the existing 128GB DGX Spark and keeps the rest of the core platform. According to the announcement (opens in a new tab), the 64GB model has the same GB10 Grace Blackwell Superchip, DGX OS and NVIDIA AI software stack as the 128GB model.

Model size is the practical difference. NVIDIA says the 64GB unit supports models of up to 100 billion parameters, while its product page lists up to 200 billion for the 128GB unit. The product page spec table gives:

• up to 1 PFLOP of FP4 tensor performance
• 64 GB or 128 GB of LPDDR5X unified memory on a 256-bit interface
• 273 GB/s memory bandwidth, with no separate figure for the 64GB model
• a ConnectX-7 NIC at 200 Gbps plus 10 GbE
• a 240 W power supply in a 150 x 150 x 50.5 mm case weighing 1.2 kg

Storage is listed as up to 4 TB, but the pages reviewed here do not say what storage each partner's 64GB unit ships with.

Clustering two units

Every DGX Spark includes a ConnectX-7 network card. Two units connected directly with a QSFP cable pool their memory to 128GB. NVIDIA says that raises the model limit to 200 billion parameters and doubles memory bandwidth. In NVIDIA's own Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7x the performance of a single system.

NVIDIA Sync Cluster Assistant handles the setup. It detects connected units, checks their configuration and configures the ConnectX-7 network. Because both nodes run the same software stack, NVIDIA says a workflow moves from one unit to two without reconfiguring the environment.

NVIDIA Sync Model Launcher is due at the end of the month. NVIDIA says it will download and start Qwen3.8 27B on one unit or a cluster, make the model reachable from a laptop, and set up OpenCode to use it for coding in a browser.

Who the 64GB DGX Spark is for

NVIDIA pitches it to developers, researchers and hobbyists who want to run agents, inference, fine-tuning or data science on their own hardware instead of renting a cloud instance for each job. The announcement gives three example workflows. One is a coding or research agent that runs around the clock. Another offloads model inference from a laptop or desktop to the Spark. The third adds a second unit when a model or context window outgrows one box.

The lower entry price is the main change for buyers. A team can start with one 64GB unit and add a second later to reach 128GB of pooled memory, rather than buying the larger configuration up front. NVIDIA does not give the price of the 128GB model in the announcement, so a cost comparison between the two paths cannot be made from these sources.

How to get started

NVIDIA's steps are short. Install a supported inference framework (it names llama.cpp, Ollama, vLLM and LM Studio), then download the recommended local model for your workflow. Our comparison of llama.cpp, Ollama and LM Studio covers the differences between the first group. To use two units, connect them through their ConnectX-7 ports and launch NVIDIA Sync Cluster Assistant.

NVIDIA says the system supports NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models, Ollama, vLLM and PyTorch with CUDA from the start. Setup guides live in the DGX Spark playbooks repository (opens in a new tab) on GitHub, which includes guides for connecting two Sparks and for serving models with vLLM. For a server you can call from code, see serving an open model with vLLM behind an OpenAI-compatible API.

Limits and open questions

Some pieces are not ready yet. NVIDIA lists three playbooks as coming soon to 64GB devices: serving LLMs with vLLM, running OpenClaw with a local LLM, and connecting multiple DGX Sparks. The Model Launcher is scheduled for the end of the month, and Blender's prebuilt installer is described only as coming soon.

The "up to 100 billion parameters" figure comes without a stated precision or quantization, so the largest model that fits in practice depends on format and context length. The 1.7x clustering result is a single NVIDIA test on one model.

Partner details are thin. The announcement gives one starting price for all six manufacturers and does not list regions, per-partner configurations or warranty terms. Dell and HP have no product links in the announcement.