TurboQuant, RotorQuant stay fork-only

This is promising, but not turnkey. If you want the newest quant tricks right now, expect to pin a fork, match model-specific flags, and tolerate breakage while the ecosystem settles.

• –Upstream llama.cpp has active TurboQuant discussion, but the working implementations are still being carried in forks/branches and described as experimental.
• –The Qwen3.6-35B-A3B-TQ3_4S card is explicit: 12.4 GiB GGUF, TurboQuant runtime fork required, and recommended launch settings use `-ctk q4_0 -ctv tq3_0 -fa on`.
• –That makes the 5060 Ti 16GB target plausible, but only if you stay within the exact build/runtime combo the model author tested.
• –For day-to-day reliability, a conventional high-quality GGUF quant on mainline llama.cpp is still the safer choice; TurboQuant is more of a bleeding-edge capacity play.