This repository is a derivative work of XiaomiMiMo/MiMo-V2.5
(https://huggingface.co/XiaomiMiMo/MiMo-V2.5\), licensed under MIT.

Every individual weight value is bit-identical to the upstream release.
This repository repacks the original tensor-parallel-sharded weights for
MLX inference on Apple Silicon:

  * The 32 source safetensors shards are concatenated into a single file.
  * Per-expert MoE weights (256 experts per layer x 3 projections) are
    stacked along a new leading axis, as required by MLX MoE loaders.
    No values change; only the storage layout (256 separate tensors -> 1
    tensor with shape [256, ...]).
  * Multimodal weights (vision encoder, audio encoder, MTP heads) are
    excluded. This repository targets the text-only inference path.
  * The config.json field model_type is changed from "mimo_v2" to
    "mimo_v2_block_fp8" to select the MLX block_fp8 model class.

No quantization, permutation, scale aliasing, or padding is applied. Anyone
with the original Xiaomi release can reproduce this file in ~10 minutes
using convert_mimo.py from this repository.

The original MiMo-V2.5 weights and architecture are the work of the Xiaomi
MiMo Team.