llama.cpp-tq3 shrinks Qwen3.5-27B, fits 16GB GPUs

This is a fit-and-efficiency win, not a universal replacement for Q4_0. The meaningful story is that 27B-class local inference just became more practical on consumer GPUs without giving up much quality.

• –The key delta is memory, not raw perplexity: about 1.5 GB saved on a 27B model can decide whether it stays entirely on GPU.
• –The approach is genuinely algorithmic, combining Walsh-Hadamard rotation, centroid quantization, and dual half-block scales instead of just repackaging existing bits.
• –The release depends on a custom llama.cpp fork, so adoption hinges on maintaining that runtime path or upstreaming the support.
• –The author’s caveats are important: this is one strong witness on one model and one card, not proof that TQ3_1S generalizes cleanly to every model size.