NVIDIA DGX Spark (128GB)

NVIDIA DGX Spark (128GB): Launched October 2025 (shown at CES 2025 as "Project DIGITS") around a single GB10 Grace Blackwell Superchip: a 20-core Arm Grace CPU (10 Cortex-X925 + 10 Cortex-A725) paired with a Blackwell GPU, sharing 128GB of LPDDR5X unified memory at 273 GB/s. The Founders Edition launched at $3,999; NVIDIA raised the MSRP to $4,699 in February 2026, citing LPDDR5X supply constraints.

Generating each reply token ("decode") on a dense model is almost entirely bandwidth-bound: this calculator estimates roughly 2.3 tok/s for a 70B dense model at Q8_0 (86 GB) on the Spark, close to the ~2.7 tok/s LMSYS independently measured for Llama 3.1 70B at FP8 (also 8-bit), since 273 GB/s is the hard ceiling regardless of inference engine. Mixture-of-experts models fare much better because decode only reads their active parameters, not the full weight set: GPT-OSS 120B (~5B active) is estimated at roughly 20 tok/s here at NVFP4, though independent llama.cpp benchmarks on real Spark hardware report 35-50 tok/s at the equivalent MXFP4 quantization; treat both as engine-dependent ballparks, not a hard ceiling. Reading and encoding the prompt before generation starts ("prefill") is a different regime: it's limited by compute, not memory bandwidth, so independent benchmarks measure 1,700-2,000+ tokens/sec chewing through long prompts on the Blackwell GPU regardless of model size.

Ships with NVIDIA DGX OS (an Ubuntu-based Linux distribution) and a preinstalled CUDA/TensorRT-LLM stack, the same software used on NVIDIA's datacenter GB200 racks, which is the Spark's core pitch: prototype locally, deploy unchanged to a cluster. llama.cpp's CUDA backend runs natively on the aarch64 Grace CPU; Ollama and vLLM both ship Spark-specific container images. Two units can be linked over the dual QSFP ports (up to 200 Gb/s aggregate) to pool memory for models too large for a single 128GB unit.