How much context fits in VRAM
Weights are only half the story: the KV cache grows with context and batch size, and it can rival the model itself at 128k. If you see an out-of-memory error after a long paste, lower the context before lowering the quant.