Grok 4.20 (Non-Reasoning)

This is the strongest throughput case, with a 260.9 tok/s median on ~2549 output tokens and no issues. The model sustains high decode speed on long, structured technical output without instability.

Performance is solid and close to the overall median at 243.3 tok/s on ~2079 output tokens. This suggests stable sustained generation under moderate-length technical prompts.

This is the main throughput weakness, falling to 155.7 tok/s median on ~1903 output tokens. The drop indicates the model slows significantly on essay-style generation, likely due to longer-form reasoning and less structured output.

TTFT is slightly better here at 1932 ms median, but still near 2 seconds even for a short 74-token response. Startup latency remains the main bottleneck rather than decode speed.

This is the slowest startup case at 2066 ms median TTFT despite only 15 output tokens. The short completion length makes the latency overhead especially visible and suggests weak first-token responsiveness.