Nvidia launches DGX Spark with 64 GB memory to enable running models and agents locally

Nvidia announced a new category of its personal supercomputer “DGX Spark” with unified memory of 64 GB, available starting Friday, October 23, at a price starting from $4,999, through its manufacturing partners: Aser, Asus, Dell, Gigabyte, HP, and MSI. This category targets researchers, software developers, and AI developers, allowing them to run autonomous agents and open models locally on their devices without the need to rely constantly on cloud infrastructure, while preserving Nvidia's software suite and the series' operating system from day one.

The integrated device relies on the high-performance GP10 Grace Blackwell chip, combined with unified memory and a ConnectX-7 networking system, and a full software acceleration stack supported by CUDA technology. These specifications support running language models up to 100 billion parameters and agent-based applications directly on the system. The stack also provides developers with a practical environment for fine-tuning, inference, data science, and system testing on proprietary data within the local work environment, freeing development teams from reserving cloud instances for each new software task.

Connecting two nodes with a networking cable doubles memory to 128 GB and raises the efficiency of processing large models.The cluster assistant application “Nvidia Sync” enables merging two units via a dedicated cable, sharing unified memory and raising total capacity to 128 GB, which expands software support to serve models up to 200 billion parameters with doubled memory bandwidth. In a test performed by the company on the “Qwen 3.8” model of 27 billion parameters, the dual-node system achieved performance about 1.7 times that of a single node, with the software automatically verifying configuration and adjusting network settings without manual intervention.

The devices ship pre-configured to support common inference frameworks, including Ollama, LLM, PyTorch, LM Studio, and LLaMA-CPP, as well as CUDA-X libraries, NEMO models, and independent agent building packages. The company plans to release a model-deployment tool by the end of the month to simplify model download and execution via the browser and integration with development environments such as Open Code, alongside support for design applications like Blender with pre-installed packages, and to provide local execution of the image-generation model “Qwen Image 2.1” on the processors.

This shift moves the building of independent agents from cloud budgets to an in-house development environment within offices.For technical engineering teams and software developers in the Gulf, Egypt, and the Arab region, this announcement changes operational cost calculations and internal system governance. Instead of draining startup project budgets by paying continuous cloud consumption subscriptions to run code-review agents or analyze sensitive documents around the clock, this device allows a one-time investment in independent local hardware. More importantly, it gives companies that adhere to strict data confidentiality or intellectual-property safeguards a practical option to keep models and customer data within the on-premises network, while preserving the flexibility to scale gradually as workloads grow.