K2 Long-term thinking model high-speed version, supports 256k context, excels at deep reasoning, output speed increased to 60-100 tokens per second.