Cydonia 24B v4.3 hits 16GB ceiling

This is the quintessential local-LLM compromise: 16GB VRAM buys you choice, not freedom. For RP, the real decision is whether you want a faster 9B model or a bigger MoE/27B setup that leans on DDR5 and accepts some offload.

• –Cydonia-24B-v4.3 Q4_K_M sits around 14.3GB as a GGUF, so it fits but leaves very little headroom once KV cache and runtime overhead enter the picture.
• –Qwen3.5 9B is the speed-first answer if you care more about tokens per second than raw model size.
• –Qwen3.5 27B Q3_K_S and Qwen3.5 35B A3B quants are the quality-first stretch options when RAM offload is acceptable.
• –KoboldCpp is a good fit for this kind of tuning because the offload, context, and GPU-layer knobs are easy to reason about.