# DGX Spark

Canonical page: https://agenticdevices.org/devices/nvidia-dgx-spark/

![NVIDIA DGX Spark](https://agenticdevices.org/images/devices/dgx-spark.png)

Product image: NVIDIA

NVIDIA / Agent computers

A compact desktop platform for running models and developing AI agents locally.

Local capable · Local AI · Developer · Desktop

See manufacturer

Configuration and seller determine pricing; no current price verified. [Source ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/)

Sources checked November 4, 2026

## Capabilities & specifications

Manufacturer specifications and disclosures. This profile is not a hands-on review or a security assessment.

| Specification | Details |
| --- | --- |
| AI capabilities | Local model inference, fine-tuning and agent development. [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |
| Processing location | Local execution supported; cloud use depends on applications and configuration. [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |
| Processor | GB10 Grace Blackwell; 20-core Arm CPU. [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |
| Sensors & inputs | Not documented in reviewed sources |
| Display & output | Not documented in reviewed sources |
| Connectivity | Wi-Fi 7, Bluetooth 5.4, 10 GbE and ConnectX-7 networking. [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |
| Memory & storage | 128 GB unified memory; 64 GB configurations through OEM partners. Up to 4 TB NVMe. [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |
| Battery or power | 240 W power supply; 140 W chip TDP. These are different measures. [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |
| Software | NVIDIA DGX OS (Ubuntu-based). [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |
| Subscription | Not documented in reviewed sources |
| Privacy & security disclosures | Local-capable hardware does not establish the data behavior of installed applications. [\[1\] ↗](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) |

## Technical notes

-   NVIDIA’s up-to-1-PFLOP figure is theoretical FP4 with sparsity, not measured agent performance.
-   Do not compare that figure directly with a device quoting INT8 TOPS.

## Sources

1.  [DGX Spark product specifications](https://www.nvidia.com/en-us/products/workstations/dgx-spark/)
    
    Manufacturer · Checked November 4, 2026
    

[Evidence policy](https://agenticdevices.org/editorial-policy/)