WIP
Unfinished EXL3 of XiaomiMiMo/MiMo-V2.6-Flash-RL. Not a release. Not validated on the GB10.
- Routed experts: target 2.25 bpw, mul1, -hq. Head: 6 bpw.
- visual., audio_encoder., speech_embeddings., and model.mtp. are full-precision copies in preserved-fp.safetensors. This build's text loader does not run those paths.
- audio_tokenizer/ and dflash/ are file copies, not quantized. The draft is DFlash, not DFlash2.
- Built to be pulled and served later on one GB10 with exllamav3.