Part 2 of 4: a benchmark journal across nixpkgs llama.cpp, upstream master, and ik_llama.cpp on Qwen3.6-27B. Six hours, four backends, all converging at 66 tok/s — and the physical reason why.
Part 3 of 4: a deep-dive into why speculative decoding silently breaks (or runs anti-economically) on hybrid attention+SSM architectures like Qwen3.6, Mamba-2, and RWKV — and what would need to change upstream to fix it.
Part 4 of 4: the actual NixOS module, llama-pull helper, claude-code-router wiring, and one-line workflow for switching models. Five Nix files for a complete, isolated, rollback-able local LLM service.
Part 1 of 4: motivation, hardware, and stack choices for serving Qwen3.6-27B locally on a 32 GB consumer GPU with NixOS, before any benchmarks or trade-offs kick in.
