NVIDIA announced DGX Spark at GTC 2026, a desktop workstation powered by the Grace Blackwell Superchip that runs AI models up to 120 billion parameters locally. The system starts at 128GB unified memory and scales to 512GB across four linked nodes, bringing enterprise-grade AI compute to a form factor that sits on a desk.

What Happened

During the GTC 2026 keynote, NVIDIA revealed DGX Spark as a desktop AI workstation built for developers, researchers, and creators who need local AI compute without relying on cloud services. The system pairs the Grace Blackwell Superchip with ConnectX-7 networking, allowing multiple units to link together for larger workloads.

A single DGX Spark handles inference and fine-tuning for models up to 120B parameters. Two linked nodes support models up to 400B parameters with 256GB total memory. At four nodes, the system runs models up to 700B parameters with 512GB memory and near-linear scaling.

Why It Matters

For creative AI practitioners, DGX Spark addresses a real gap between consumer GPUs and cloud-based compute. Running a 120B parameter model locally means image generation, video synthesis, and large language model inference can happen on your desk without per-token API costs or data leaving your network. The ability to fine-tune models locally is especially valuable for creators building custom workflows with proprietary training data.

The multi-node scaling makes DGX Spark practical for small studios. Two desks, two units, and a cable give you 400B parameter capacity. That is enough to run the largest open-source models locally without any cloud dependency.

Key Details

  • Processor: NVIDIA Grace Blackwell Superchip (Grace CPU + Blackwell GPU).
  • Memory: 128GB unified per node, expandable to 256GB (2 nodes) or 512GB (4 nodes).
  • Networking: ConnectX-7 NICs with RoCE 200 GbE for multi-node configurations.
  • Performance: Qwen 3 Coder 80B at 2,390 tokens/second (32K context). Nemotron 3 Super 120B at 2,855 tokens/second (128K context).
  • Framework support: TensorRT LLM, vLLM, SGLang, OpenClaw agent framework.
  • Developer tools: Tile IR and cuTile Python enable code portability between DGX Spark and cloud Blackwell GPUs.

What to Do Next

Creators and developers interested in local AI compute should review the DGX Spark technical overview on the NVIDIA Developer blog. The system is designed for those who want to move from cloud prototyping to local production. Pricing and availability details are expected ahead of the fall 2026 launch window alongside the Vera Rubin platform.