DGX Spark 64GB: Who NVIDIA's $4,999 Local AI Box Is For

NVIDIA announced a 64GB version of DGX Spark on October 2, 2026, giving local AI builders a cheaper way into its Grace Blackwell desktop platform. The new configuration keeps the same GB10 superchip and software stack as the 128GB original, starts at $4,999, and ships from partner brands on Friday, October 23.

• Price: from $4,999, sold through Acer, ASUS, Dell, Gigabyte, HP and MSI.
• Availability: Friday, October 23, 2026.
• Capacity: NVIDIA says one unit runs models up to 100 billion parameters, and two clustered units handle up to 200 billion.
• New software: NVIDIA Sync Cluster Assistant now, with a one-click Model Launcher due at the end of October.

What Is the DGX Spark 64GB?

DGX Spark is NVIDIA's compact desktop AI system built around the GB10 Grace Blackwell Superchip. The 64GB model halves the unified memory of the original but, according to NVIDIA, keeps DGX OS, the full NVIDIA AI software stack and ConnectX-7 networking. Tom's Hardware and The Register both report that memory bandwidth stays at 273 GB/s and the 20-core Arm CPU is unchanged, so the main difference is how large a model fits in memory at once.

The 128GB DGX Spark remains on sale for people working with larger models or fine-tuning, per Tom's Hardware.

How Big a Model Can 64GB Run?

NVIDIA rates the 64GB unit for models up to 100 billion parameters on the device. That covers a wide range of today's open-weight models, including mid-size coding and reasoning models running on quantized weights. NVIDIA's own showcase example is Qwen3.8 27B, which fits with plenty of room left for long context.

If you outgrow one box, two 64GB units joined with a QSFP cable pool their memory to 128GB and support models up to 200 billion parameters. NVIDIA says the pair delivered up to 1.7x the performance of a single system in its Qwen 3.8 27B testing. That is a vendor figure, so treat it as a best-case reference. The Register notes ConnectX-7 supports clusters of up to four units.

What Does NVIDIA Sync Add?

Clustering used to mean manual network setup. The new NVIDIA Sync Cluster Assistant detects connected units, validates the configuration and sets up the ConnectX-7 link automatically. Later this month, NVIDIA Sync Model Launcher is due to download and launch Qwen3.8 27B on a single unit or a cluster with a click, and Tom's Hardware reports it will also set up the OpenCode browser-based coding agent.

Out of the box, NVIDIA lists support for Ollama, vLLM, PyTorch with CUDA, NVIDIA Agent Toolkit and Nemotron models, with a Blender installer coming soon. Anyone who already runs a local LLM with Ollama on a mini PC will find familiar tools on much more capable hardware.

Who Should Buy the 64GB Model?

The sweet spot is developers, researchers and small teams who want a private, always-on AI workstation for agents, coding assistants and retrieval projects without renting cloud GPUs. If your models fit under 100 billion parameters, the 64GB DGX Spark gives you the same software experience as its bigger sibling at a lower entry price.

For a different approach to local LLM hardware, compare it with the x86-based 192GB GMKtec EVO-X5 Pro, or revisit how NVIDIA framed multi-unit setups in our DGX Station and Spark clustering coverage. More compact AI machines live in our mini PC section.

Sources: NVIDIA Blog — October 2, 2026; Tom's Hardware — October 2, 2026; The Register — October 2, 2026.