Qwen3.8-27B on 24GB VRAM: 131k Context with MTP Enabled

sisyphus-cycle · reddit · 2026-08-17

A user shared their llama.cpp configuration for running Qwen3.8-27B (Q4KM) on a 24GB VRAM GPU (RTX 4090). The testing reveals a trade-off between context length and generation speed controlled by MTP (Multi-Token Prediction). Without MTP, the model achieves 194k context tokens at 40 tps. With MTP enabled, context drops to 131k tokens, but generation speed increases to 65 tps (peaking at 80 tps for Python code).

Original post →

More from Infra

Infra channel →