GLM-5

This is the weakest sustained-throughput case at 67.2 tok/s median, likely reflecting the longer 2952-token generations and more complex instruction-following load. It is still stable with no failures, but decode speed drops below the model's overall median.

Throughput is near the overall median at 73.3 tok/s, indicating consistent decode performance on medium-length technical output. No issues were recorded, so this prompt looks representative of the model's baseline speed.

This is the fastest sustained-throughput prompt at 77.0 tok/s median, helped by the shortest average output length among throughput tasks. The result suggests the model maintains or slightly improves decode rate when generation length is lower.

Startup latency is strong here at 711 ms median TTFT, the best TTFT result in the set. The very short 54-token output likely reduces prefill and generation overhead, making this the model's fastest response-start case.

This is the slowest startup case at 1501 ms median TTFT despite only 12 average output tokens, which points to prompt-side prefill or retrieval-like overhead rather than decode cost. The gap versus ttft-definition is large enough to suggest TTFT is sensitive to prompt content, not just output length.