October 2, 2026
Run local models on the new DGX Spark: starting at $4,999
DGX Spark with 64 GB of memory goes on sale on Oct 23, 2026 through six NVIDIA partners, starting at $4,999.
NVIDIA is adding a version with 64 GB of unified memory alongside the 128 GB DGX Spark. The company says it supports running models with up to 100 billion parameters locally.
Same platform. NVIDIA has kept the GB10 Grace Blackwell chip, DGX OS and its AI software stack. The new version will be sold exclusively through partners:
- Acer - ASUS - Dell - Gigabyte - HP - MSI
You can work with Spark from your usual laptop. NVIDIA Sync supports Windows, macOS and Debian/Ubuntu: after installation, add the device on the same network by name or IP address and enter its credentials.
Two devices together. NVIDIA says a pair of 64 GB Spark devices provides 128 GB of combined memory and supports models with up to 200 billion parameters. In the company's test with Qwen3.8 27B, the pair ran up to 1.7 times faster than a single device.
Pairing requires a ConnectX-7 connection using a single QSFP cable and both devices added to Sync. Start setup through Settings → Cluster Assistant → Add New Cluster. You will need SSH, sudo privileges and system software from April 2026 or later.
Cluster Assistant configures the network; running models and fine-tuning are separate steps covered in NVIDIA's instructions. The vLLM recipe recommends Qwen3.8-27B NVFP4 for one or two Spark devices. With a pair, the model is split across the devices, while DFlash2 speeds up response generation by proposing tokens in advance.
The local AI agent NemoClaw is installed with `curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash`. Express Install on Spark offers a model, local inference through vLLM and an isolated agent environment. The installer adds Node.js 22.16+ if it is missing.
NVIDIA is still adapting the instructions for vLLM, OpenClaw and distributed tasks to the 64 GB version.
Original source: [NVIDIA on DGX Spark 64GB, Oct 2, 2026](https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync).
By the end of October 2026, NVIDIA promises Model Launcher in Sync for downloading and running Qwen3.8 27B on a single Spark or a cluster, and configuring OpenCode for coding in the browser.
