# NOTICE — sources and credits

This is an independent community build. It is **not** an official Xiaomi release.

| Component | Source | License |
|---|---|---|
| Base model | [XiaomiMiMo/MiMo-V2.6-Flash-MOPD](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD) | MIT |
| Importance matrix (computed on MiMo-V2.6-Flash-RL, routers identical), cut to the kept experts | [Baekpica/MiMo-V2.6-Flash-RL-Mixed-Quant-GGUF](https://huggingface.co/Baekpica/MiMo-V2.6-Flash-RL-Mixed-Quant-GGUF) | MIT (calibration data under their own terms, see that repo) |
| Expert-usage reference for the selection (full RL model) | [TrevorJS/MiMo-V2.6-Flash-RL-GGUF](https://huggingface.co/TrevorJS/MiMo-V2.6-Flash-RL-GGUF) v2.1 (used only for measurements) | MIT |
| REAP saliency bands (REAP25 ⊂ REAP37 ⊂ REAP50 pruned-expert lists) | [tacodevs MiMo-V2.6-Flash-RL MLX REAP releases](https://huggingface.co/tacodevs) (`compression_alloc.json`) | MIT |
| Decode-trace hot-expert list | [SSHdotCodes/mimo-v2.6-flash-pro6000](https://github.com/SSHdotCodes/mimo-v2.6-flash-pro6000) (`profiles/hot-experts.txt`) | MIT |
| REAP method | [CerebrasResearch/reap](https://github.com/CerebrasResearch/reap), arXiv:2510.13999 | Apache-2.0 (method only, no code used) |
| Runtime | [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) + [ek15072809/Swap-MoE](https://github.com/ek15072809/Swap-MoE) (SSD expert streaming) | MIT |

Own contributions: routing measurements of the full model (German text, code, and the model's own on-policy German answers), the combined expert-selection rule, the ranged downloader/converter from the original weights, the quantization recipe, the runtime patches in runtime/patches (expert lock, UTF-8 sanitizing, DFlash for MiMo, Swap-MoE GPU guard) and all evaluations reported in the model card.