The new Kimi 3 model features 2.8 trillion parameters and natively visual capabilities, positioning it as a strong performer. It utilizes Kimi Delta attention to efficiently manage KV cache, addressing a key memory bottleneck in large language models.