Tag
Antirez reports benchmarking DS4 inference on the DGX Spark (GB10), noting 12 tokens/sec generation speed and high prefill performance, with plans to merge the codebase once mature.