Wësche(@WescheNex1q):Reproduced @u1tra_instinct new full GLM-5.3 TP4 image with Tensorfold on our own 4× DGX Spark Two runs: prose: 40.7 / 43.4 tok/s (his 41.8) code: 36.6 / 40.6 tok/s (his 38.1) 25K-token prefill + needle: 22.7–24.0 s (his ~25 s), PASS clocks 2.4–2.5 GHz, 68–71 °C, 0 swap Long reasoning: 27.1 → 35.4 tok/s vs the previous image.

Wësche

@WescheNex1q

Day time artist and night time AI enthusiast. Building & benchmarking frontier LLMs on 4x DGX Spark clusters + Mac. Creator of Vesica Studio. Houston

加入 January 2013

731 正在關注 3.1K 粉絲

Wësche@WescheNex1q

2026.10.05 03:50

Reproduced @u1tra_instinct new full GLM-5.3 TP4 image with Tensorfold on our own 4× DGX Spark Two runs: prose: 40.7 / 43.4 tok/s (his 41.8) code: 36.6 / 40.6 tok/s (his 38.1) 25K-token prefill + needle: 22.7–24.0 s (his ~25 s), PASS clocks 2.4–2.5 GHz, 68–71 °C, 0 swap Long reasoning: 27.1 → 35.4 tok/s vs the previous image.