We publish benchmarks only from runs we can document (hardware, software version, model file, settings). DGX Spark results: TBD.
Meanwhile, the theoretical decode ceiling from its 273 GB/s bandwidth for a Q4_K_M 70B dense model (≈43 GB of weights) is about 6.3 tokens/s; for gpt-oss-120b (5.1B active) it is far higher. Ceilings are not benchmarks. See all benchmarks.