NVIDIA launches DGX Spark 64GB, a 1-petaFLOP Grace Blackwell desktop for local agents

DGX Spark is NVIDIA's personal AI computer line, and the new configuration adds 64GB of LPDDR5x unified memory around the GB10 Grace Blackwell superchip. The machine delivers 1 petaFLOP of FP4 compute. FP4 is the low-precision format Blackwell uses to speed up inference, and 64GB of unified memory lets developers run mid-sized open-weight models on a desk. Two units can be linked for 128GB of coherent memory at 273 GB/s, enough for larger models or fine-tuning runs that would not fit on one box.

The target uses are local agent token generation, fine-tuning and inference. The economic pitch is avoiding recurring cloud API fees. For agents that run around the clock and generate millions of tokens, a one-time hardware purchase can beat per-token pricing. Dell, ASUS and HP are partners, so the hardware will ship through standard enterprise channels. Meta's Muse Glimmer, distilled from Muse Spark, is also available as an NVIDIA NIM, scoring 51.2 on SWE-Bench Pro with a 131K context. But full BF16 Glimmer needs 55GB or more, leaving little headroom on a single 64GB unit.

The launch fits a week of momentum for local AI. Amazon's Strands Decider 2B runs in milliseconds on a consumer GPU, Kolibri-1 activates only 3.46B parameters, and a practitioner ran a 125B Qwen MoE on a single RTX 3090 Ti by offloading to system RAM. Apple's high-memory Macs are the obvious competitor for local inference. Separately, F5 benchmarks showed NVIDIA DPU routing tripling AI throughput at peak GPU memory load, though only for NIM deployments.

The caveats are memory bandwidth and price. 273 GB/s is modest next to datacenter GPUs, so token generation speed on large dense models will lag. The sources also omit pricing, which will decide whether the box beats renting cloud GPUs. Watch for real tokens-per-second numbers on popular open models and how the 64GB version compares with high-end Macs.