Local LLM sizing gets practical

The post is a very common local-LLM pain point: model selection is still too manual, and hardware-fit calculators are filling that gap.

• –8GB VRAM is usually enough for smaller quantized dense models, but longer context windows can eat the saved memory fast.
• –Community advice leans toward 7B-9B class models first, then MoE or offload-friendly models if system RAM is strong enough.
• –Tools like `llmfit` and `willitrunai.com` are useful because they encode the messy fit/performance tradeoff instead of forcing users to do the math by hand.
• –Runtime details matter: Windows, LM Studio, quantization choice, and CPU offload can swing tokens/sec more than raw parameter count.