Xiaomi releases MiMo-V2.5-DFlash model weights
Xiaomi has uploaded the official MiMo-V2.5-DFlash to Hugging Face. DFlash appears intended to speed up text generation from MiMo-V2.5, a model with more than 300 billion parameters. One firsthand setup reportedly runs the existing model at about 8–10 using two 24GB s and memory offloading to a system with 96GB or 128GB of DDR5 memory.
DFlash might double that speed, but no measured result was provided. Xiaomi also released a separate MTP model, which predicts several upcoming tokens to generate text faster. The current llama.cpp cannot use the previously released MTP head because it has trouble recognizing its layers.
The separate MTP model or DFlash might avoid that problem, but both possibilities remain untested. Someone still needs to convert the weights to GGUF and check , speed, and output quality.
Key points
- Official MiMo-V2.5-DFlash weights are available on Hugging Face.
- The base model has more than 300 billion parameters.
- One setup reportedly produces about 8–10 with two 24GB s and extra .
- The possible twofold speed increase from DFlash has not been measured yet.
- GGUF conversion and testing are still needed.