Claude Sonnet 4.5 sparks local model hunt

There isn’t a real local Sonnet 4.5 clone here; on a 16GB card, the win is getting close enough for RP, not matching frontier behavior exactly.

• –Qwen3.5 27B and Ministral-3-14B look like the sweet spot on this rig because they can be quantized and offloaded to the GPU.
• –The uncensored or “abliterated” suggestions matter because RP users care as much about tone and refusal behavior as raw intelligence.
• –A 70B-class model might feel closer to Sonnet in style, but it is a rough fit for 16GB VRAM unless you accept big compromises.
• –The fact that the OP later remembered KoboldCpp is a reminder that the local stack, not just the checkpoint, shapes results.
• –This thread is really about cost: users want frontier-like outputs without API spend, and local models still have a ceiling.