# NVIDIA DGX Spark 64GB: Build and Run AI on Your Desktop > Markdown version of an Informer Tech article. Canonical HTML page: https://www.informertech.com/post/nvidia-dgx-spark-64gb-local-ai > Source: Informer Tech, https://www.informertech.com/ - **Canonical URL:** https://www.informertech.com/post/nvidia-dgx-spark-64gb-local-ai - **Publisher:** Informer Tech (https://www.informertech.com/) - **Category:** Hardware Reviews, AI Industry News - **Author:** Tomáš Král - **Published:** 2026-11-01 - **Description:** Explore the NVIDIA DGX Spark 64GB desktop AI system. See GB10 Grace Blackwell specs, unified memory, and how to fine-tune and run local AI models. --- NVIDIA DGX Spark 64GB brings model fine-tuning, inference, and autonomous AI agents to a compact desktop system. It combines the GB10 Grace Blackwell Superchip, 64GB of coherent unified memory, and NVIDIA’s AI software stack, with built-in networking for connecting additional Spark systems as workloads grow. The 64GB configuration is offered exclusively through Acer, ASUS, Dell, Gigabyte, HP, and MSI. NVIDIA positions it for developers, researchers, students, data scientists, and AI enthusiasts who want to build and test local models. ## Fine-Tune AI Models With Your Own Code and Data DGX Spark supports a workflow built around changing a model, testing it, and refining the results. Instead of sending every iteration to an external inference service, developers can work with the model and inspect its output locally. One use case NVIDIA identifies is fine-tuning an existing large language model with a developer’s own source code. The resulting model can then run locally for future development tasks, allowing the developer to explore how well it understands that codebase’s conventions. The system also supports computer-vision projects, data-science work, and local inference testing. These workloads let developers move between experimentation and testing within the same local environment. That flexibility matters when a project develops beyond a single prompt. An experiment might begin with a private document collection or a software repository, then grow into an agent that uses tools and completes several steps. DGX Spark is designed to support that progression. ## DGX Spark 64GB Hardware Specifications The GB10 Grace Blackwell Superchip pairs a Blackwell GPU with a 20-core Arm CPU. Through NVLink-C2C, the CPU and GPU use a coherent architecture that gives both access to the same memory pool. NVIDIA rates the system for up to 1 petaFLOP of FP4 AI performance, alongside 273GB/s of memory bandwidth.
**Component** **Specification**
Superchip NVIDIA GB10 Grace Blackwell
GPU Blackwell GPU
CPU 20-core Arm CPU
Memory 64GB of coherent unified memory
CPU–GPU connection NVLink-C2C
AI performance Up to 1 petaFLOP at FP4
Memory bandwidth 273GB/s
Cluster networking Built-in NVIDIA ConnectX-7
Operating system DGX OS
The design centers on the compute and memory requirements of local AI. A conventional desktop can support many experiments, but a larger model, longer context, and supporting application workloads place additional demands on the system. ## Unified Memory for Models and Larger Workloads A local AI application needs memory for more than its main model. Longer contexts, additional models, and supporting workloads also need space as the application becomes more complex. DGX Spark’s 64GB memory pool is intended to accommodate those demands. NVIDIA highlights newer open models, including: - Qwen3.8-27B - Meta Muse Glimmer - NVIDIA Nemotron 3.5 Lightning NVIDIA describes these models as fitting within a 64GB memory footprint while offering capabilities comparable to frontier models from a few months earlier. Additional memory capacity also creates room for richer context and supporting workloads alongside the main model. When a project exceeds a single system’s capacity, two 64GB units can combine their memory into a 128GB pool through ConnectX-7 networking and NVIDIA Sync Cluster Assistant. ## DGX OS and Local AI Software DGX Spark runs DGX OS and comes with NVIDIA’s CUDA-accelerated AI software stack. Included tools such as PyTorch, Jupyter, and Ollama support prototyping, fine-tuning, and inference. Having those tools available is intended to reduce the setup work involved in starting a local AI project. Developers can use the system to modify a model, run an experiment, examine the results, and repeat the process. The local workflow also gives builders control over the models, data, and environment used in their experiments. NVIDIA positions DGX Spark as a way to keep models, prompts, data, and inference on the device without per-token cloud inference fees. Cloud services still have a role in projects that require enormous amounts of compute. DGX Spark focuses on the local experimentation and development work that can happen before a project reaches that scale. ## Build Autonomous Agents and Multi-Agent Workflows Agent workloads involve more than generating a single response. A coding agent might examine a project, find a problem, change the code, test that change, and continue based on the result. A research workflow could involve several specialist models, tool calls, and context that remains available throughout the task. NVIDIA describes DGX Spark as a platform for these demands, including long-context reasoning, multi-agent pipelines, and local inference. Multiple models and agents can use the system’s shared memory pool. NVIDIA OpenShell adds policy-based privacy and security controls for agent workloads. Smaller projects are also part of the intended use. An enthusiast could build an agent to work through a personal dataset, check a software project, organize information, or carry out a repetitive sequence of tasks. The local environment gives the builder control over how those experiments run. ## Connect Multiple DGX Spark Systems DGX Spark includes ConnectX-7 networking so a local setup can expand beyond one machine. Connecting two 64GB systems provides a combined 128GB memory pool and up to 2 PFLOPS of total AI compute. NVIDIA says that, in the configuration described in its brief, this pair can deliver up to 1.7 times the performance of a DGX Spark 128GB system. That figure applies to the specified configuration. NVIDIA also describes clusters containing up to four DGX Spark systems. These configurations provide larger shared memory pools for models and workloads that exceed one unit’s capacity, with technical material covering distributed inference and fine-tuning across multiple systems. ### Cluster Setup With NVIDIA Sync NVIDIA Sync helps manage the connected systems through its Cluster Assistant. Its functions include discovering devices, configuring networking, routing workloads, and monitoring system health. The aim is to reduce the amount of cluster administration developers need to handle as their projects grow. Additional Spark systems can extend the local environment without requiring a completely different workflow. ## Desktop Size and Configuration Options DGX Spark measures 150 × 150 × 50.5mm and weighs 1.2kg. NVIDIA describes it as power-efficient enough to operate from a standard wall outlet, supporting workloads that need to remain active between workstation sessions. The 64GB model’s partner availability was scheduled for October 23. It is supplied exclusively through Acer, ASUS, Dell, Gigabyte, HP, and MSI. For users who need more memory in a single system, NVIDIA also offers the DGX Spark 128GB configuration through NVIDIA and its OEM partners.