Comment by lacoolj

This is cool. I would like to try using this model with Qwen3.8 or Gemma4 as a speculative decoding model. Wonder if that would help give the overall results with the higher quality output