Which statement is safe to use when planning memory for a quantized local model?

Which statement is safe to use when planning memory for a quantized local model?

1. Raw quantized weight size is a lower bound; measure buffers and the KV cache at the target context and concurrency.
2. A 4-bit model always needs exactly one quarter of the memory of its 16-bit version.
3. If the model file is smaller than VRAM, every supported context length is safe.