Xiaomi uploads MiMo-V2.5-DFlash weights and separate MTP model to Hugging Face

Xiaomi has uploaded the official DFlash weights for its MiMo-V2.5-DFlash model, along with a separate MTP (Multi-Token Prediction) model, to Hugging Face.

• The repository includes a dedicated `dflash` directory containing the model weights.
• A separate MTP model is also available, addressing previous issues where Llama.cpp struggled to identify MTP layers.
• The source notes that DFlash could potentially double inference speed compared to standard execution on consumer hardware.

The availability of these official weights allows users to experiment with GGUF conversions and test the compatibility of the DFlash and separate MTP implementations.