BetterBench results

	Date	Model	Runtime	GPUs	GPU	Decode t/s	Prefill @64k	Usability

Single-stream, batch = 1

Category	Runs	TTFT p50 ms	ITL 1% low	ITL med	ITL 99% high	Decode med	IQR	CV

Concurrency sweep

Level	Ok	Aggregate t/s	TTFT p50 ms	TTFT p99 ms	Decode/req t/s

Prefill sweep

Depth tok	Prompt tok med	TTFT p50 ms	1% low	Median	99% high