or Qwen 3.5 122B A10B, both use more memory and still have experts sized for decent speed at the 395’s memory bandwidth at 4bit quantization