RX 7900 XT Powers Qwen3-30B-A3B

Hot take: this is a reasonable “buy the GPU first” situation, but it is not a magic fix for the whole machine.

• –The RX 7900 XT’s 20GB VRAM is the key spec here; that is what makes larger quantized local models realistically usable.
• –The Ryzen 5600G is not the bottleneck for inference in most local-LLM setups; it is fine as an older host CPU.
• –Qwen3-30B-A3B is a MoE model with 3.3B activated parameters, so it is more approachable than a dense 30B model, especially when quantized.
• –16GB system RAM is workable but tight, especially if you want longer contexts, multiple apps open, or CPU offload.
• –On AMD, the big question is software support: Linux + ROCm or llama.cpp-style workflows are usually the safer path than assuming everything will “just work.”
• –If the goal is coding help rather than just “largest model possible,” you may get a better day-to-day experience from a smaller, faster model with less memory pressure.