19 DGX Spark recipes for current models benchmarked in one table

RigBase compiled Mia's AI Lab open-source DGX Spark recipes into a single table of decode and prefill figures. GLM-5.3 Flash (320B) hits 60.4 tok/s on two Sparks and 95.2 on four, Qwen3.8 Flash-Next runs at 48.7 tok/s on one, and full GLM-5.3 (753B) needs three Sparks for 31.3 tok/s.

• GLM-5.3 Flash on 2/3/4 Sparks: 60.4 / 77.6 / 95.2 tok/s, prefill up to 2,583 tok/s
• Qwen3.8 Flash-Next on a single Spark: 48.7 tok/s in NVFP4 via vLLM
• DeepSeek V4.1 Flash: 74.9 tok/s on two Sparks, 196.6B params kept in Engram tables
• Full GLM-5.3 (753B) on three Sparks at 2.75 bits: 31.3 tok/s