Through positional encoding improvements such as RoPE and ALiBi, sliding window attention, and long-text fine-tuning, the effective context windows of mainstream models have been extended to the scale of hundreds of thousands to even millions of tokens.