Qwen3.8 27B
on NVIDIA RTX 5080 · llama.cpp · 85,000 ctx
Sep 7, 2026
Summary
User reports hybrid quantization to fit MTP and long context in 16 GB VRAM.
Setup uses MTP draft with spec-type draft-mtp and spec-draft-n-max 3, at 85,000 context, with q4_0 KV cache quantization for both cache types.