Throughput, latency and KV cache headroom for Qwen3.8-27B measured on a single RTX PRO 6000, comparing BF16, FP8 and a serving baseline across context lengths and batch token settings.