kernelpool/MiMo-V2.6-Pro-RL-MXFP4-GGUF

MiMo-V2.6 Pro RL GGUF for DwarfStar

GGUF files for the MiMo-V2.6 Pro RL port in DwarfStar (DS4), branch kernelpool/ds4:mimo-v26 (pull request: TODO link). Metal only. The model runs over tensor parallelism between two 512 GB Macs.

Files

The main GGUF is above the Hugging Face single-file limit, so it is stored in two parts and joined after download, the same way the official DwarfStar Q4 release is.

File	Bytes	SHA-256
MiMo-V2.6-Pro-RL-MXFP4.gguf.part1	480,000,000,000	2c53578328dedbb3462f01422fa0548eccd5567ff8e2ab7254aebf74d0c3aea9
MiMo-V2.6-Pro-RL-MXFP4.gguf.part2	77,146,510,560	601b0ad28e8e36bf88adabb97e4994ffc9da281ec26488671350645850ed969e
MiMo-V2.6-Pro-RL-MXFP4.gguf (joined)	557,146,510,560	dac336f68cecf88b8ca04ab4ea2c75b831752e0323c7152a758082826f60bf21
MiMo-V2.6-Pro-RL-DFlash-Q8_0.gguf	2,941,587,648	3aec4c9819dacad46e5bca5ccd55b92e426e807616783f8140914854e76e4e91

The main file uses the mimo2 layout: the checkpoint's own MXFP4 experts repacked without requantization, Q8_0 attention, dense and output weights, BF16 embeddings, and the three MTP blocks. The DFlash sidecar uses the dflash layout plus the mask embedding and value scale DS4 needs.

hf download kernelpool/MiMo-V2.6-Pro-RL-MXFP4-GGUF --local-dir gguf
cd gguf
cat MiMo-V2.6-Pro-RL-MXFP4.gguf.part2 >> MiMo-V2.6-Pro-RL-MXFP4.gguf.part1
mv MiMo-V2.6-Pro-RL-MXFP4.gguf.part1 MiMo-V2.6-Pro-RL-MXFP4.gguf
rm MiMo-V2.6-Pro-RL-MXFP4.gguf.part2
openssl dgst -sha256 MiMo-V2.6-Pro-RL-MXFP4.gguf   # dac336f6...

Appending the second part to the first keeps the extra disk space to the size of the second part.

Run

Both Macs need the joined file locally. Each rank keeps half of the attention heads, half of the routed experts and half of the vocabulary head, about 270 GiB resident, so each Mac needs 512 GB. The link setup is in docs/DISTRIBUTED.md. Start the worker first:

git clone -b mimo-v26 https://github.com/kernelpool/ds4 && cd ds4 && make
# Machine B (worker)
./ds4 -m gguf/MiMo-V2.6-Pro-RL-MXFP4.gguf --ctx 8192 --mtp \
  --tensor-parallel --role worker --coordinator 192.168.0.1 9911 --transport rdma
# Machine A (coordinator)
./ds4-server -m gguf/MiMo-V2.6-Pro-RL-MXFP4.gguf --ctx 8192 --mtp \
  --tensor-parallel --role coordinator --listen 192.168.0.1 9911 --transport rdma

--mtp drafts with the MTP blocks inside the main file; both ranks run the draft and verify it together, so pass it to both or to neither. Greedy output follows plain decoding.

The server exposes mimo-v2.6-pro, mimo-v2.6-pro-chat (thinking off) and mimo-v2.6-pro-reasoner (thinking on).

DS4 does not run the DFlash sidecar or vision under tensor parallelism yet, so the sidecar is not used for now and no vision encoder is included. See docs/MIMO_V26.md.

Conversion

Written by gguf-tools/mimo26_quantize.py from XiaomiMiMo/MiMo-V2.6-Pro-RL at revision 54b10491b1811c76aa9681a9d0ff872396a4064c. 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 over two-Mac tensor parallelism on 100 official continuations from the Xiaomi platform with the fixture in gguf-tools/quality-testing/mimo-v2.6-pro-20260923; the reference scores are in the repository's QA_BEFORE_RELEASES.md.

The original checkpoint is released by Xiaomi under the MIT license, which applies to these files as well.