NVIDIA announced the DGX Spark 64GB on October 2, 2026, providing a new hardware configuration for local AI development. The system features 64GB of unified memory and a starting price of $4,999, arriving from manufacturer partners like Dell and HP on October 23. (unified memory: a shared memory architecture where the central processing unit and graphics processing unit access the same memory pool.) If you currently rely on cloud GPU instances for model inference, this hardware provides an on-premise alternative that removes that dependency. (model inference: the process of using a trained machine learning model to make predictions or generate outputs from new input data.)
Is this for you?
If you develop local AI, the DGX Spark 64GB offers an on-premise alternative to cloud GPU instances starting at $4,999.
- If you rely on cloud GPU instances for model inference, this hardware provides an on-premise alternative that removes that dependency.
- If you are an AI developer building local agents, the Spark provides a production-ready environment that runs agentic AI privately.
- If you are a research scientist working with open weights, you can use the 64GB or 128GB memory options to scale hardware.
- Skip this if you are a mobile app developer as the DGX Spark is a workstation-class platform.
- Skip this if you only want software-only AI tools as this is a hardware-centric announcement.
What you can do now
- If you plan to purchase from Dell or HP, note the starting price of $4,999.
- If you intend to scale to 200 billion parameters, plan to purchase two DGX Spark 64GB units to pool memory to 128GB.
Key figures
- Unified memory capacity
- 64GB
- Starting price
- $4,999
- Max model support
- 200 billion parameters (clustered)
- Availability date
- Friday, Oct. 23
- Primary compute
- GB10 Grace Blackwell Superchip
The $4,999 DGX Spark 64GB hardware architecture
The DGX Spark 64GB configuration utilizes the GB10 Grace Blackwell Superchip to provide a complete local AI platform. It combines this compute power with unified memory—a shared memory architecture where the CPU and GPU access the same memory pool—and NVIDIA ConnectX-7 networking. This hardware setup supports up to 100-billion-parameter models on a single unit.
NVIDIA says the platform retains the same GB10 Grace Blackwell Superchip and DGX OS
NVIDIA says the platform retains the same GB10 Grace Blackwell Superchip and DGX OS found in the 128GB model. The 64GB variant serves as a more accessible entry point for local AI development. It provides a space to experiment with models and private data without constant cloud access.
Scaling to 200 billion parameters via clustering
Two DGX Spark 64GB units can cluster together to pool memory to 128GB and expand model support to up to 200 billion parameters. This scaling happens via the NVIDIA Sync Cluster Assistant, which detects connected units and configures the ConnectX-7 network automatically. Users can connect the units directly using a QSFP cable without additional configuration hurdles. (QSFP cable: a high-speed physical connection used to link networking equipment and high-performance computing components.)
Clustering provides significant performance
Clustering provides significant performance gains for specific models. NVIDIA says two clustered 64 GB systems delivered up to 1.7x performance compared with a single system in its Qwen 3.8 27B test. The platform is designed for a growth path where users can start small and scale as workloads demand.
Software support for Ollama and vLLM
The DGX Spark ships with the NVIDIA AI software stack, including the NVIDIA Agent Toolkit and CUDA-X AI libraries. It supports popular runtimes like Ollama, vLLM, and PyTorch with CUDA out of the box. Developers can go from power-on to running models in minutes.
NVIDIA will release the NVIDIA Sync Model Launcher at the end of the month
NVIDIA will release the NVIDIA Sync Model Launcher at the end of the month to simplify the launch process. This tool allows users to download and launch models like Qwen3.8 27B with a few clicks. The launcher will also set up OpenCode to allow developers to code in a browser using the local model.
Who this is for: AI Developers and Researchers
AI developers building local agents should use the Spark because it provides a production-ready environment that runs agentic AI—systems designed to perform multi-step tasks as autonomous agents—privately on device. Research scientists working with open weights can use the 64GB and 128GB memory options to scale their hardware based on the specific size of models like Qwen or Nemotron. (open weights: machine learning models where the internal parameters are publicly available for anyone to download and use.)
System integrators and specifiers should note that the platform is available through major partners like Dell and HP starting at $4,999. Creative professionals using Blender can also utilize the platform, as Blender is one of the first creator applications to support the hardware with a prebuilt installer.
Who this is not for: Mobile App Developers
Mobile app developers should skip this announcement because the DGX Spark is a workstation-class platform, not mobile hardware. These users should instead look for mobile-specific inference optimization guides or edge-device hardware. Users only interested in software-only AI tools should also ignore this hardware-centric announcement and stick to standard cloud-based AI service documentation.
A walkthrough for Zara’s local AI setup
Zara runs a small software firm and needs to deploy a coding agent that analyzes sensitive client documents locally. She buys one NVIDIA DGX Spark 64GB for $4,999 and uses the pre-installed Ollama runtime to deploy her agent. When her project grows to require a model that exceeds 64GB of memory, she buys a second unit and connects them via a QSFP cable. The NVIDIA Sync Cluster Assistant automatically pools the memory to 128GB, allowing her to run larger models without changing her software environment.
If Zara did not require data privacy for her clients, she would continue renting cloud GPU time. Cloud instances provide a different scaling path but lack the ‘always-on’ local accessibility of the Spark.
Applications for the DGX Spark 64GB
| Use Case | Who it serves |
|---|---|
| Running 24/7 coding or research agents | Software Engineers |
| Offloading heavy inference from a laptop | Content Creators |
| Fine-tuning models on private data | Privacy-conscious Researchers |
| Running Qwen3.8 27B locally | Developers |
| 3D modeling with AI-ready hardware | 3D Artists |
What this replaces for cloud users
The DGX Spark reduces the need for cloud instances for every task, removing the recurring costs and privacy risks of remote hosting. This allows developers to maintain a constant local environment for model experimentation and inference.
What is not yet known: NVIDIA did not specify the exact shipping dates for the specific
NVIDIA did not specify the exact shipping dates for the specific manufacturer SKUs. The company also did not disclose the specific power consumption or thermal requirements for the 64GB configuration. The performance gains for models larger than 200 billion parameters remain unconfirmed.
NVIDIA DGX Spark 64GB: questions the original report leaves open
- Who made the decision or is responsible?
- The original report did not say who made the decision.
- Why did it happen?
- The original report did not say why it happened.
Reported by NVIDIA on 2 October 2026: NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI.