NVIDIA said DGX Spark systems with 64GB of unified memory will be available starting Friday, Oct. 23, from Acer, ASUS, Dell, Gigabyte, HP and MSI at $4,999. The compact systems include DGX OS and NVIDIA’s AI software stack, and can run local AI agents and models on device without relying on a cloud connection. NVIDIA says the new configuration supports models with up to 100 billion parameters and retains the GB10 Grace Blackwell Superchip found in the 128GB model.

DGX Spark combines Grace Blackwell compute, ConnectX-7 networking and CUDA-accelerated software for local agents, inference, fine-tuning, data science and edge development. Developers can experiment with models and their own data on the system. NVIDIA lists support from launch for Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and Ollama, vLLM and PyTorch with CUDA; it says models can be running within minutes. Blender is also preparing a downloadable installer for the platform.

Two 64GB systems can connect directly over a QSFP cable and pool memory to 128GB, supporting models up to 200 billion parameters. NVIDIA says its Qwen 3.8 27B test found up to 1.7x performance from two clustered systems compared with one, alongside twice the memory bandwidth. NVIDIA Sync Cluster Assistant detects connected devices, validates their configuration and sets up the ConnectX-7 network, allowing the same software environment to scale without reconfiguration.

NVIDIA Sync Model Launcher, scheduled for the end of the month, will let users launch Qwen3.8 27B on one system or a cluster and make it accessible from laptops. It will also configure OpenCode for browser-based coding with the model. Example uses include keeping a coding or research agent running continuously, running language or image generation on DGX Spark while a PC handles other tasks, and connecting two systems when a workload exceeds one unit’s capacity. NVIDIA’s setup instructions name llama.cpp, Ollama, vLLM and LM Studio as supported inference frameworks. The article also points to forthcoming playbooks for serving LLMs with vLLM, running OpenClaw with a local LLM, and connecting multiple DGX Sparks for distributed workloads.

NVIDIA said DGX Spark systems with 64GB of unified memory will be available starting Friday, Oct. 23, from Acer, ASUS, Dell, Gigabyte, HP and MSI at $4,999. The compact systems include DGX OS and NVIDIA’s AI software stack, and can run local AI agents and models on device without relying on a cloud connection. NVIDIA says the new configuration supports models with up to 100 billion parameters and retains the GB10 Grace Blackwell Superchip found in the 128GB model.

DGX Spark combines Grace Blackwell compute, ConnectX-7 networking and CUDA-accelerated software for local agents, inference, fine-tuning, data science and edge development. Developers can experiment with models and their own data on the system. NVIDIA lists support from launch for Agent Toolkit, CUDA-X AI libraries, Nemotron open models, and Ollama, vLLM and PyTorch with CUDA; it says models can be running within minutes. Blender is also preparing a downloadable installer for the platform.

Two 64GB systems can connect directly over a QSFP cable and pool memory to 128GB, supporting models up to 200 billion parameters. NVIDIA says its Qwen 3.8 27B test found up to 1.7x performance from two clustered systems compared with one, alongside twice the memory bandwidth. NVIDIA Sync Cluster Assistant detects connected devices, validates their configuration and sets up the ConnectX-7 network, allowing the same software environment to scale without reconfiguration.

NVIDIA Sync Model Launcher, scheduled for the end of the month, will let users launch Qwen3.8 27B on one system or a cluster and make it accessible from laptops. It will also configure OpenCode for browser-based coding with the model. Example uses include keeping a coding or research agent running continuously, running language or image generation on DGX Spark while a PC handles other tasks, and connecting two systems when a workload exceeds one unit’s capacity. NVIDIA’s setup instructions name llama.cpp, Ollama, vLLM and LM Studio as supported inference frameworks. The article also points to forthcoming playbooks for serving LLMs with vLLM, running OpenClaw with a local LLM, and connecting multiple DGX Sparks for distributed workloads.