license: mit
language:

• en
• zh
  base_model:
• dealignai/MiMo-V2.6-Flash-RL-UNCENSORED
  base_model_relation: quantized
  tags:
• gguf
• llama.cpp
• mimo2
• mixture-of-experts
• multimodal
• uncensored
• experimental

MiMo-V2.6-Flash-RL-UNCENSORED — IQ2_XXS with IQ2_S down projections

An experimental, aggressively quantized GGUF build of
dealignai/MiMo-V2.6-Flash-RL-UNCENSORED.
It is intended for local llama.cpp inference when the full MXFP4 GGUF is
impractical. This is a quality-versus-memory trade-off, not a lossless
conversion.

Provenance

Quantization

The build was produced with llama-quantize, an importance matrix, and
--keep-split. The default quantization is IQ2_XXS, except for the 47
expert FFN down-projection tensors, which use IQ2_S. Attention tensors and
the first/final/MTP blocks are kept as Q8_0; token embeddings use Q5_K.

Quality check

This comparison uses llama-perplexity on 16 sequential 512-token chunks of
WikiText-2 wiki.test.raw, with the source MXFP4 GGUF as the reference. It is
a small reproducible signal, not a full capability or safety evaluation.

Relative to MXFP4, this build raises perplexity by 37.4% and disagrees
with its top-probability token on 23.87% of the evaluated positions. It
improves on the all-IQ2_XXS build, but remains an aggressive quantization. Do
not use it where maximum fidelity is required without task-specific testing.

Files and checksums

Usage

Use a recent llama.cpp build with support for the required GGUF tensor
types. Keep all four split files in the same directory and point to the first
part:

llama-cli -m .\MiMo-V2.6-Flash-RL-UNCENSORED-IQ2_XXS-DOWN-IQ2_S-00001-of-00004.gguf `
  -ngl 0 -c 8192 --chat-template-kwargs '{"enable_thinking":false}'

The inherited chat template supports enable_thinking both on and off. Start
with a modest context and increase it only after checking memory use on your
hardware. -ngl 0 is only a conservative CPU example, not a performance
recommendation.

Validation performed

• All 508 output tensors were present with shapes matching the source.
• Four output shards were hashed with SHA-256 (above).
• A local smoke generation answered 42 to “What is 17 + 25?”.
• The WikiText-2 comparison above completed without pipeline errors.

Limitations and safety

This quantization inherits the behavior, capabilities, and limitations of its
upstream model. The upstream model is explicitly labelled UNCENSORED;
this conversion does not add safety training, moderation, or suitability
guarantees. Validate outputs for your intended use, especially for high-impact
or safety-sensitive applications.

Acknowledgements

Thanks to the maintainers of the upstream model, the source GGUF conversion,
and llama.cpp.