Running Qwen 3.8 27B full-context on a 12GB laptop: two llama.cpp configs

PyaesoneP · reddit · 2026-08-19

After buying a gaming laptop (RTX 5070 Ti Mobile 12GB, Core Ultra 9 275HX, 32GB DDR5), the author found it GPU-poor for local LLMs—but still got Unsloth Qwen 3.8 27B UD Q4KXL running via llama.cpp + CUDA with full context and q80 KV cache, at 1.5–5 t/s. They share two tuned configs: (1) an agentic coding setup with OpenCode, -ctx 262144, FFN layers offloaded to CPU, flash attention, q8 KV cache, and MTP speculative decoding—slow (1.5 t/s at 180K context) but with enough context to hold the big picture; it's the first local model they trust on real (personal) projects. (2) A speed-prioritized personal assistant config with Hermes Agent (9–11.5 t/s, 98K context) that also fixed instruction-following on SOUL.md/MEMORY.md/USER.md rules the previous MoE model frequently ignored. Tensor offloading, credited to community posts, made a surprisingly big difference.

Original post →

More from coding & agent

coding & agent channel →