NVIDIA Announces DGX Spark 64GB: A 1-PetaFLOP Grace Blackwell Desktop for Local AI Agents, Fine-Tuning, and Inference

Jean-marc Mommessin at MarkTechPost reports NVIDIA announced a 64GB DGX Spark configuration from Acer, ASUS, Dell, Gigabyte, HP and MSI. Same GB10 Grace Blackwell superchip as the original — up to 1 petaFLOP FP4 with sparsity, 20-core Grace Arm CPU, ConnectX-7 — but 64GB unified LPDDR5x instead of 128GB. Availability: October 23, 2026 via NVIDIA Marketplace, OEMs and retail.

NVIDIA positions 64GB as enough for capable 30–35B-class open models; two units cluster to 128GB memory and more compute. The piece cites token consumption up 14x since early 2026 as agents burn tool calls, retries and long context — costs that vanish on owned hardware. Software on first boot includes DGX OS, PyTorch, Jupyter, Ollama, plus NemoClaw, OpenShell guardrails and Nemotron optimizations. Wall-outlet power, 1.2 kg, no server room.

We’re for putting serious inference on a desk. Always-on agents on metered APIs turn “helpful” into a line item. A Grace Blackwell mini that runs quantized Muse Glimmer, Nemotron 3.5 Lightning or Qwen3.8-27B locally — then clusters when you outgrow one box — is the hardware answer to that bill.

Useful AI should be runnable near the data and the human. Unified CPU/GPU memory over NVLink-C2C matters for agents that keep several models, KV caches and tool processes alive. Shipping Ollama and policy guardrails on day one is how you make that real for developers who aren’t cluster admins.

The downside if desktop AI factories proliferate is energy, noise and a new class of always-on local agents with file and network reach — privacy wins versus cloud, security responsibility shifts to the owner. 64GB also isn’t magic: full BF16 Glimmer needs 55GB+ and leaves almost nothing for context; quantized builds are the practical path. DeepSeek V4 Flash still wants multiple boxes in NVIDIA’s own framing.

I’m watching street pricing from OEM partners, real agent workloads on one 64GB unit versus the 128GB SKU, and whether clustering stays a one-click story or a weekend project.

MarkTechPost lists example local models: Meta’s Muse Glimmer (~17GB quantized), NVIDIA Nemotron 3.5 Lightning (NVFP4), and Qwen3.8-27B (~13.5GB at 4-bit), with KV and runtime overhead on top.