spiritfather/MiMo-V2.6-Flash-MOPD-Yamz-Uncensored-GGUF

MiMo-V2.6-Flash-MOPD-Yamz-Uncensored — GGUF

🧬 Ablation: yamz-labs/MiMo-V2.6-Flash-MOPD-EXL3-Yamz-Uncensored · base: XiaomiMiMo/MiMo-V2.6-Flash-MOPD
📦 Source GGUF: QuantaPlanta/MiMo-V2.6-Flash-MOPD-MXFP4-GGUF
🛠️ Runs on llama.cpp (master)
📊 Benchmarked on CaliperBench
💬 Discord

All credit for the uncensoring goes to yamz-labs, who fitted the refusal direction and per-layer strengths for MiMo-V2.6-Flash-MOPD and published them as uncensor_direction.st + uncensor_spec.json. Credit for the base model goes to XiaomiMiMo (309B total / ~15B active MoE, 256 experts, 8 active), and for the lossless GGUF conversion these quants start from to QuantaPlanta. These are quantizations only.

Why this repo exists. yamz-labs' uncensor is applied at runtime by their Kyojin engine on an EXL3 pack (tested on AMD Strix Halo); stock ExLlamaV3, TabbyAPI, llama.cpp, LM Studio and Ollama ignore it and serve the base model with its refusals. This repo bakes the same edit into the weights so it works in any llama.cpp-based runtime — Apple Silicon, CUDA, CPU — with no special engine or flag.

Benchmarked on CaliperBench — a creative-writing benchmark scoring prose craft, roleplay and willingness rather than general intelligence.

Refusals

100 held-out harmful prompts (harmful_behaviors eval split), thinking on, greedy, 1024-token cap, classified by CaliperBench's refusal detector. Same prompts and settings for every row.

Model	Refused	Notes
Stock MiMo-V2.6-Flash-MOPD (QuantaPlanta MXFP4)	97 / 100
Attention-only bake (experts left stock)	74 / 100	what you get if the expert half of the edit is lost
This repo, Q4_K_M	3 / 100	0 empty or truncated answers
yamz-labs EXL3 + Kyojin runtime edit (their card)	4 / 100	their prompt set and regex detector, not directly comparable

The edit is one global direction: it lowers refusals and over-refusals together. It is not a safety evaluation and says nothing about what the model will or will not produce.

Provided quants

KLD, PPL and same-top-token are measured against the stock model (QuantaPlanta's MXFP4 GGUF, which is lossless against the release) on wiki.test.raw, 100 chunks at ctx 512, llama-perplexity --kl-divergence. So each number is the combined cost of the uncensor edit and the quantization, which is what you actually pay versus running stock. The table fills in as each size lands.

Quant	Size	Mixture (default / gate exps / up exps / down exps)	PPL	vs stock	KLD vs stock	Same top token
Q3_K_M	141.2 GiB	Q8_0 / Q3_K / Q3_K / Q4_K+Q5_K	5.3146 ± 0.0745	+3.65%	0.0635 ± 0.0009	89.57%
IQ4_XS	158.4 GiB	Q8_0 / IQ4_XS / IQ4_XS / IQ4_XS+Q5_K	5.2126 ± 0.0726	+1.66%	0.0333 ± 0.0004	91.81%
Q4_K_M	176.5 GiB	Q8_0 / Q4_K / Q4_K / Q4_K+Q6_K	5.1724 ± 0.0724	+0.88%	0.0243 ± 0.0003	93.26%
Q5_K_M	206.5 GiB	Q8_0 / Q5_K / Q5_K / Q5_K+Q6_K	5.1632 ± 0.0722	+0.70%	0.0173 ± 0.0002	94.22%
Q6_K	238.4 GiB	Q8_0 / Q6_K / Q6_K / Q6_K	5.1515 ± 0.0720	+0.47%	0.0146 ± 0.0002	94.67%
stock (reference)	162.9 GiB	BF16 / MXFP4 / MXFP4 / MXFP4	5.1275 ± 0.0714	0	0	100%

Loading (any current llama.cpp)

llama-server -m MiMo-V2.6-Flash-MOPD-Yamz-Uncensored.Q4_K_M.gguf -ngl 99 -fa on -c 32768 -ctk q8_0 -ctv q8_0 --jinja

Thinking: on by default in MiMo's template. For non-thinking replies pass chat_template_kwargs: {"enable_thinking": false} per request.

Sampling: yamz-labs and Xiaomi recommend temperature 1.0, top-p 0.95; greedy decoding is not recommended for generation.

How it was made

1. Source. QuantaPlanta/MiMo-V2.6-Flash-MOPD-MXFP4-GGUF: a stock convert_hf_to_gguf.py of the release — bf16 attention/dense, routed experts in MXFP4. MiMo ships its experts natively in MXFP4, so this file is lossless against the original weights.

2. Bake the edit. yamz-labs' spec states the equivalent weight form of their runtime edit, W ← W − w(L)·r̂ r̂ᵀ W, applied to the writers into the residual stream: attn_output (layers 11–47) and the MLP/MoE output (ffn_down in dense layer 0, routed ffn_down_exps in layers 1–46), with their per-layer strengths. MTP layers, router, gate/up experts, norms and embeddings are untouched. Math is float32; the edited tensors are written as BF16 into a master GGUF (~320 GB), and every edited tensor was checked against the expected |1 − w| shrink of the direction component (e.g. layer 20 experts 0.222× vs 0.221× expected).

   • The edited experts are deliberately not re-encoded to MXFP4: that erased ~80% of the expert edit, and that build refused 74/100.

3. Quantize each size from that master with llama-quantize: routed experts take the nominal type; every non-expert tensor (attention, dense and MTP FFN, nextn.eh_proj, token embeddings, output head) is pinned to Q8_0; router, norms and attention sinks stay F32. Static quants, no imatrix. --allow-requantize is needed only because the unedited gate/up experts are MXFP4 in the master — the model's native precision, so this is still a first quantization.

   llama-quantize --allow-requantize \
     --tensor-type attn_qkv=q8_0 --tensor-type attn_output=q8_0 \
     --tensor-type 'ffn_(gate|up|down)\.weight=q8_0' --tensor-type nextn.eh_proj=q8_0 \
     --token-embedding-type q8_0 --output-tensor-type q8_0 \
     MASTER.gguf MiMo-V2.6-Flash-MOPD-Yamz-Uncensored.Q4_K_M.gguf Q4_K_M
   

Built with llama.cpp master 4e416ee (stock mimo2 support, no patches).

Use

The edit lowers refusals. You decide how to use the model and you answer for it under the laws that apply to you. Intended uses: research, evaluation, red-teaming, creative writing and private deployment. Do not use it to produce illegal content or to target people, or in a public service without your own filtering and moderation.

License

MIT, inherited from XiaomiMiMo/MiMo-V2.6-Flash-MOPD.