Modelpublic

Baekpica/MiMo-V2.6-Flash-RL-GGUF

sourceHugging Facemitupdated 5d agoView on Hugging Face

1likes590downloads

Model Card

MiMo-V2.6-Flash-RL GGUF — calibration reference

This repository holds intermediate artifacts actually generated while building MiMo-V2.6-Flash-RL Mixed-Quant GGUF. The final compact mixed variant belongs in that separate repository.

The source is XiaomiMiMo/MiMo-V2.6-Flash-RL, pinned to `3b38d063180c3e4aed9691fdc735f3d10b266ee4`.

Reference representation

The four MXFP4-BF16 shards preserve the original routed experts through an exact MXFP4 repack and expand source FP8 dense matrices to BF16. Control tensors are stored as F32. This is not a full-BF16 source checkpoint or a Q8_0 baseline. It provides original-checkpoint values for importance-matrix collection without recalibrating from the final IQ2 weights.

Shard	Bytes
`MiMo-V2.6-Flash-RL-MXFP4-BF16-00001-of-00004.gguf`	44,499,128,352
`MiMo-V2.6-Flash-RL-MXFP4-BF16-00002-of-00004.gguf`	44,493,179,840
`MiMo-V2.6-Flash-RL-MXFP4-BF16-00003-of-00004.gguf`	44,493,179,840
`MiMo-V2.6-Flash-RL-MXFP4-BF16-00004-of-00004.gguf`	41,422,766,592
Total	174,908,254,624

Keep all shards together and open the first shard. The language artifact includes the checkpoint's three embedded MTP blocks; this is not evidence of validated speculative decoding. Multimodal encoders and the separate DFlash model are separate components and are not supplied by these four shards alone. They are included as the auxiliary files listed below.

Auxiliary artifacts

File	Role	Validation
`mmproj-MiMo-V2.6-Flash-RL-BF16.gguf`	BF16 image/audio input encoders and projector, about 2.75 GB	One-image native/GGUF numerical comparison; full runtime media validation pending
`MiMo-V2.6-Flash-RL-DFlash-Q8_0.gguf`	Five-layer Q8_0 draft, about 1.56 GB	63-tensor mapping/shape audit, exact F32 and sampled Q8 error checks; runtime acceptance untested
`dflash/`	Original config, learned mask embedding, upstream example, and explicit runtime contract	Source files retained byte for byte

The DFlash GGUF explicitly preserves 64 rotary dimensions per 128-dimensional head and attention value scale 0.612. Its config, sinks, target-layer mapping, and learned mask embedding must be consumed by a compatible runtime. The bundled upstream Python example does not implement every MiMo-specific draft setting; the runtime contract records the production reference. No stock-runtime compatibility is claimed.

The projector records the production MiMo image pixel bounds (8,192–8,388,608), alongside production ImageNet normalization. The pixel-bound correction preserved all 811 tensor payloads; see projector-bounds-audit.json. Both temporal patch weight slices match the source exactly (video-patch-weight-audit.json).

Validation status

—All 90 downloaded source repository files passed Hub checksum verification.
—Independent MXFP4 repacking and tensor-parallel QKV ordering checks passed.
—All 36,096 expert matrices were covered by a deterministic sampled-row audit: 108,257 rows matched the independently repacked source. See reference-audit.json; this is not an exhaustive payload comparison.
—The reference loaded on a B300 GPU and passed a short arithmetic decode check.
—Original-representation text imatrix collection is running as of 2026-09-22.
—Final mixed-model quality, multimodal end-to-end behavior, MTP/DFlash execution, and DGX Spark serving remain pending. No throughput or benchmark qualification is claimed here.

artifact-manifest.json records exact file sizes and SHA-256 digests. SHA256SUMS can be checked after download:

bash

hf download Baekpica/MiMo-V2.6-Flash-RL-GGUF --local-dir ./MiMo-reference
cd ./MiMo-reference
sha256sum -c SHA256SUMS

Chat template and reproduction

chat_template.jinja is copied byte for byte from the pinned source. Use the original MiMo tokenizer and special-token mapping. Correct image, audio, and video processing additionally requires the corresponding native encoder and input protocol.

The conversion uses a local MXFP4 extension to llama.cpp revision 5836771. Reproduction scripts will accompany the mixed release and private Spark handoff. The reference's approximately 175 GB file size is not the compact DGX Spark target; see the separate mixed model card for that recipe and its current estimates.

License

MIT, inherited from the pinned upstream model.