NVIDIA launches 64GB DGX Spark for local AI models up to 100 billion parameters

NVIDIA introduced a 64GB unified-memory configuration of DGX Spark for local AI development and inference. Sales through manufacturer partners begin Friday, October 23, at $4,999.

The new version retains the GB10 Grace Blackwell chip, DGX OS and NVIDIA’s full software stack. It supports local models with up to 100 billion parameters without requiring a cloud connection.

Acer, ASUS, Dell, Gigabyte, HP and MSI will offer the system. It includes NVIDIA Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and support for Ollama, vLLM and PyTorch with CUDA.

Two systems combine for 128GB

Two DGX Spark units can be connected with a QSFP cable through their built-in NVIDIA ConnectX-7 network adapters. NVIDIA Sync Cluster Assistant configures the network and workload distribution, pooling memory to 128GB and extending model support to 200 billion parameters.

In NVIDIA’s Qwen 3.8 27B test, a two-system cluster delivered up to 1.7 times the performance of a single unit. NVIDIA also said the configuration doubles memory bandwidth.

Software for local agents

NVIDIA plans to release Sync Model Launcher at the end of the month. It is designed to launch Qwen3.8 27B on one DGX Spark or a cluster and configure OpenCode to use the model in a browser.

Blender will be among the first major creator applications to support the platform, with a prebuilt installer planned for a later release. NVIDIA also cited uses including always-on coding and research agents, image generation, and moving inference from an everyday PC to DGX Spark.

Users can work with llama.cpp, Ollama, vLLM or LM Studio. Scaling requires connecting two systems through their ConnectX-7 ports and launching NVIDIA Sync Cluster Assistant.