NVIDIA Announces DGX Spark 64GB: A 1-PetaFLOP Grace Blackwell Desktop for Local AI Agents, Fine-Tuning, and Inference
MarkTechPost Jean-marc Mommessin ● Covered by 2 sources
NVIDIA added a 64GB DGX Spark desktop for local AI work. It’s built for agents and fine-tuning without paying per token.
Based on reporting by MarkTechPost, Jean-marc Mommessin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
NVIDIA has a new 64GB DGX Spark configuration, and the pitch is simple: keep more AI work on a desktop instead of pushing every token through a metered cloud API. The system comes from Acer, ASUS, Dell, Gigabyte, HP, and MSI, all using NVIDIA’s GB10 Grace Blackwell platform.
The smaller model keeps the same core ingredients as the original Spark: the GB10 superchip, NVIDIA’s CUDA AI software stack, and ConnectX-7 networking. The change is memory. Instead of 128GB, this version uses 64GB of unified LPDDR5x, which NVIDIA says is enough for today’s 30–35B-class open models. The box still advertises up to 1 petaFLOP of FP4 AI compute with sparsity.
That unified memory design matters more than the spec sheet gloss. CPU and GPU share one pool over NVLink-C2C, so there’s no shuffling weights back and forth between system RAM and VRAM. For agent work, that means models, KV caches, and tool processes can live together in one address space. NVIDIA is also shipping DGX OS, based on Ubuntu, with PyTorch, Jupyter, and Ollama ready at first boot. NemoClaw installs with a single command, and OpenShell adds policy guardrails.
The real target here is not one giant chatbot. It’s always-on agents, local fine-tuning, and same-day model evaluation. NVIDIA points to models like Muse Glimmer, Nemotron 3.5 Lightning, and Qwen3.8-27B as practical fits, while larger workloads can move to the existing 128GB Spark. If one unit is not enough, two 64GB systems can cluster through ConnectX-7 for 128GB of pooled memory and more compute. NVIDIA says that setup can reach up to 1.7x the performance of a single 128GB Spark.
Availability is set for October 23, 2026 through NVIDIA Marketplace, OEM partners, and retail. The hardware is small enough to sit on a desk and ordinary enough to run from a wall outlet, which is basically the point: make local AI feel less like a lab project and more like something people can leave on overnight.
My take — AI-written commentary, not fact-checked reporting
This is NVIDIA doing what NVIDIA does best: taking a real pain point, then selling the cure in a very expensive box. But the push toward local agents is sensible, because paying by the token turns automation into a subscription trap with extra steps. The boring truth is that more AI work should stay off someone else’s meter.
Read more about this at: MarkTechPost