Planned RTX 3090 rung

Qwen3.6 27B is the next single-3090 intelligence/cost test.

The hypothesis is simple: the 27B dense Qwen base may be a better coding-agent brain than the 35B-A3B MoE Ornith path, even if we need a lower quant or a tighter runtime profile to fit it on 24GB.

Current Hypothesis

Start with Huihui-Qwen3.6-27B-abliterated-MTP-GGUF. The known public candidate sizes make the 3090 lane interesting: Q4 around 15.7 GiB, Q5 around 18.2 GiB, Q6 around 20.9 GiB, and Q8 around 27.1 GiB. Q8 is not a single-3090 target at useful context; Q5 or Q6 plus careful KV/runtime work is the likely search.

This page is intentionally a plan until the profile is measured. The Ornith page is the current single-3090 result.

Experiment Questions

  • Can Qwen3.6 27B keep a useful quant and at least 64K context on one 3090?
  • Does Q5 beat the Ornith Q5 profile on coding-agent quality at similar cost?
  • Is Q6 feasible with acceptable KV cache and wall-clock speed?
  • Does MTP help enough in this runtime to matter for agent work?
  • What is the cheapest profile that stays above 5 wall tokens per second?

Baseline To Beat

The current 3090 bar is Ornith 35B Q5 at 128K: about 69.7 server tokens per second and 64.2 wall tokens per second at ctx122880/out8192, with quality evidence still unresolved. Qwen does not need to beat that speed. It needs to clear the speed floor and look smarter.