I'm using Qwen3.6 27B Q4, max context with pi on 32GB VRAM (although I'm testing out Glimmer on a feature implementation literally right now). Pi is great because it has minimal context added by the agent.

Looking forward to the 3.8 27B release to compare.