GLM-5.2-Int4-Int8 achieves ~1,200 t/s prefill and 33–54 t/s decode on 8× GB10

The article reports benchmark results for the GLM-5.2-Int4-Int8 model running on eight GB10 accelerators. The configuration achieves approximately 1,200 tokens per second during prefill and an average of 33–54 tokens per second during decoding.

• Prefill throughput reaches ~1,200 t/s.
• Average decode throughput ranges from 33 to 54 t/s.
• The setup supports tensor parallelism (tp) 8 for both GLM-5.2 and a parallel Mimo 2.5 model handling image/audio input.

This performance profile allows the system to run GLM-5.2 alongside other models like Mimo 2.5 on the same hardware.