LocalLLaMA weighs RTX 3060 model picks

This is less a product launch than a snapshot of where local AI is today: 12GB VRAM is enough to do useful work, but not enough to ignore quantization, context length, and offload strategy.

• –The thread reinforces the common 12GB rule of thumb: 7B to 8B models are the safe default, while larger models need careful quantization or RAM offload
• –Qwen-family models keep coming up because they tend to balance instruction following, coding, and general usefulness well on consumer hardware
• –The IT/sysadmin use case matters: users typically want stronger retrieval, troubleshooting, and structured reasoning than pure chat benchmarks reflect
• –Advice like “use llmfit” shows the community is converging on fit calculators and benchmarks instead of guessing from parameter counts alone
• –A few outlier claims about running much larger models on a 3060 should be treated skeptically unless the setup is explicitly documented