More GLM 753B testing on 4 DGX Sparks Four people chatting at once: Full GLM-5.3, TensorFold TP4: -4 streams × 3,000 tokens, thinking on 12,000 tokens in 194.6 s 61.7 tok/s aggregate · ~16 tok/s each First token in under 1 s for all four. His 118 tok/s is real: we reproduced it with his benchmark (113–136 greedy). We just tested a heavier workload: 4 different prompts, sampled, 3K-token thinking-on replies. 1 stream: ~35–43 tok/s Setup: @u1tra_instinct 2026-10-04-opt image, TensorFold TP4, 32K context per stream, built-in MTP drafts (no DFlash2), temp 0.7, seeds 1–4, max_tokens 3000. Recorded live in Chrome. His published 118.5 tok/s is greedy, 256-token replies with a different script; his sampled 4-stream run is 69–71.