Qwen3.8-27B 105-200 tokens per second on a 5090? If you’re running agents or tool calls, cut it in half before you plan anything around it. Anything that only happens because of Multi-Token Prediction. 💀 Single-token decode of Qwen3.8-27B Q4_K_M on a 5090 is hard-capped at 105 tok/s by memory bandwidth. Higher numbers only appear when Multi-Token Prediction is enabled. MTP drafts multiple tokens per weight read. The output stays identical, but the bandwidth math changes. Acceptance rate is the real variable: - Prose to 0.71 to big speed-up - Tool calls/JSON to 0.34 to much smaller speed-up Based on this 24 GB card is : 51 t/s (no MTP) to 128 t/s (MTP prose) to 68 t/s (MTP agent) 200 tok/s is a prose benchmark. For agentic workloads, treat it as roughly half.