Decode · tok/s
35
Prefill
—
tok/100W
14.6
Full GLM-5.3 (753B) on 4x DGX Spark. Context 64K → 256K. Sparkdash: prose c1 32.7 → 35.0 tok/s, code 37.9 → 42.0. FP4 KV + NVFP4 weights for attention/shared/MTP.
First-person 4× Spark update. Rank is the quoted Sparkdash prose 35.0. Code 42.0 stays in the note. Separate from the Int4/Int8 30.0 prose pair and the EXL3 3.0bpw 38.4 code pair.
