---

license: mit

base_model: XiaomiMiMo/MiMo-V2.6-Flash-RL

tags:

• gguf
• mimo
• metal

---

MiMo-V2.6 Flash GGUF for DwarfStar

GGUF files for the MiMo-V2.6 Flash port in DwarfStar (DS4),

branch kernelpool/ds4:mimo-v26

(pull request: TODO link). Metal only.

| File | Size | Content |

| --- | ---: | --- |

| MiMo-V2.6-Flash-MXFP4.gguf | 157.4 GiB | Main model, mimo2 layout: the checkpoint's own MXFP4 experts repacked without requantization, Q8_0 attention, dense and output weights, BF16 embeddings, the three MTP blocks |

| MiMo-V2.6-Flash-DFlash-Q8_0.gguf | 1.5 GiB | DFlash drafter sidecar, dflash layout plus the mask embedding and value scale DS4 needs |

| MiMo-V2.6-Flash-Vision-F32.gguf | 2.7 GiB | Vision encoder, lossless (recommended) |

| MiMo-V2.6-Flash-Vision-Q8_0.gguf | 0.7 GiB | Vision encoder, Q8_0 matrices, smaller and measurably less exact |

Run

git clone -b mimo-v26 https://github.com/kernelpool/ds4 && cd ds4 && make
./ds4 -m MiMo-V2.6-Flash-MXFP4.gguf --mtp
./ds4 -m MiMo-V2.6-Flash-MXFP4.gguf --mtp --vision MiMo-V2.6-Flash-Vision-F32.gguf
./ds4-server -m MiMo-V2.6-Flash-MXFP4.gguf --mtp --vision MiMo-V2.6-Flash-Vision-F32.gguf

--mtp drafts with the MTP blocks inside the main file and is the faster

option; --mtp-model MiMo-V2.6-Flash-DFlash-Q8_0.gguf uses the DFlash

sidecar instead. Both verify against the target, so temperature-zero output

follows plain decoding. The main file needs a 192 GB or larger Mac.

See docs/MIMO_V26.md.

Conversion

Written by gguf-tools/mimo26_quantize.py and gguf-tools/mimo26_vision.py

from XiaomiMiMo/MiMo-V2.6-Flash-RL at revision

3b38d063180c3e4aed9691fdc735f3d10b266ee4. The converter de-interleaves the

tensor-parallel chunks of the fused QKV projection and keeps the released

MXFP4 expert blocks bit for bit; the source revision is recorded in the

GGUF metadata.

Quality

Scored on 100 official continuations from the Xiaomi platform with the

fixture in gguf-tools/quality-testing/mimo-v2.6-flash-20260922; the

reference scores are in the repository's QA_BEFORE_RELEASES.md. The vision

encoder reproduces the checkpoint's tower on identical inputs.

The original checkpoint is released by Xiaomi under the MIT license, which

applies to these files as well.