llama.cpp rig eyes 56GB VRAM model picks

This is the right kind of overbuilt local setup: once you have 56GB of VRAM, the game shifts from “what fits” to “what gives the best quality, context, and speed tradeoff.”

• –56GB makes 30B-class models the easy default and keeps 70B-class models in play if you pick the right quantization and context settings.
• –llama.cpp’s GGUF workflow and `-hf` support make it easy to swap between models, so the real test is less about one perfect pick and more about benchmarking a few serious candidates.
• –Qwen3-family models are a sensible starting point here, especially for coding and mixed reasoning use cases.
• –If the goal is fun rather than pure text quality, multimodal GGUFs are a better way to spend spare VRAM than chasing ever-bigger chat models.
• –The interesting part of this thread is not the workstation itself, but the point it reaches: local inference gets much more compelling once you can experiment above the 30B tier.