JetBrains introduced Mellum 2, advertising latency comparable to Qwen2.5-Coder 7B achieved through Multi-Token Prediction (MTP). However, the published GGUF files do not contain layers resembling an MTP head.

  • The source questions how to extract MTP weights directly from the provided safetensors files.
  • Links are provided to a specific Q8_0 GGUF and the main model repository on Hugging Face.