spiritfather/MiniMax-M3-uncensored-GGUF

MiniMax-M3-uncensored — GGUF (imatrix quants)

<p align="center"> 🧬 Source model: <a href="https://huggingface.co/ressl/MiniMax-M3-uncensored" target="blank">ressl/MiniMax-M3-uncensored</a> · base: <a href="https://huggingface.co/MiniMaxAI/MiniMax-M3" target="blank">MiniMaxAI/MiniMax-M3</a> <br> 🧮 imatrix + recipe: <a href="https://huggingface.co/AesSedai/MiniMax-M3-GGUF" target="blank">AesSedai/MiniMax-M3-GGUF</a> <br> 🛠️ Runs on <a href="https://github.com/ggml-org/llama.cpp" target="blank">llama.cpp</a> (master) <br> 📊 Scored on <a href="https://caliperbench.com/m/minimax-m3-uncensored/" target="blank">CaliperBench</a> <br> 💬 <a href="https://discord.gg/rhE9zdKyE4" target="blank">Discord</a> </p>

[!Note] All credit for the model goes to [ressl](https://huggingface.co/ressl) — abliteration (refusal removal) of o_proj and every down_proj across all 60 layers, routers untouched — and [MiniMaxAI](https://huggingface.co/MiniMaxAI) for the base MiniMax-M3 (427B total / 10B active MoE, 128 experts, 4 active). These are quantizations only.

[!Important] The importance matrix and the per-tensor quant recipe are [AesSedai](https://huggingface.co/AesSedai)'s, taken verbatim from AesSedai/MiniMax-M3-GGUF (bartowski calibration_datav5, 49 chunks). Each size here is tensor-for-tensor type-matched to their stock MiniMax-M3 quant of the same name, so the only difference from their files is the uncensored weights. Thank you.

[!Tip] Benchmarked on [CaliperBench](https://caliperbench.com) — a creative-writing benchmark scoring prose craft, roleplay and willingness rather than general intelligence. This build is read against the stock MiniMax-M3 served from the same AesSedai recipe.

imatrix (importance matrix) quants spend their bits where the calibration text says the model is sensitive; below Q5 they are usually preferable to a static quant of the same size. Files over 50 GB are split into shards with llama-gguf-split — point llama.cpp at shard 00001 and it loads the whole set. Text-only; unpruned (all 128 experts kept).

Provided quants

PPL, the PPL-ratio column and KLD are measured against this model's own Q8_0 on wiki.test.raw (512-token chunks, llama-perplexity --no-repack), so they read as quantization loss of the uncensored weights, not a comparison to the stock model. 1-(Mean PPL(Q)/PPL(base)) is the PPL excess over that Q80 base (lower is better; Q80 is the reference at 0). The table fills in as each size lands.

Loading (any current llama.cpp)

llama-server -m IQ4_XS/MiniMax-M3-uncensored-IQ4_XS-00001-of-00005.gguf -ngl 99 -fa on -c 32768 -ctk q8_0 -ctv q8_0 --jinja

The IQ4_XS fits a 256 GB Apple Silicon box.

Thinking: on by default in MiniMax-M3's template. For the non-thinking mode pass chat_template_kwargs: {"thinking_mode": "disabled"} per request.

Quant recipe

1. 1.ressl/MiniMax-M3-uncensored bf16 safetensors → Q8_0 GGUF with convert_hf_to_gguf.py (llama.cpp b10819).
2. 2.Each smaller size cut from that Q80 with `llama-quantize --allow-requantize` using AesSedai's `imatrix.gguf` and their mixture: the ftype sets the dense tensors and `--tensor-type` overrides push only the routed experts down. Example for IQ4XS:

   llama-quantize --allow-requantize --imatrix imatrix.gguf \
     --tensor-type ffn_gate_exps=iq3_s --tensor-type ffn_up_exps=iq3_s --tensor-type ffn_down_exps=iq4_xs \
     MiniMax-M3-uncensored-Q8_0.gguf MiniMax-M3-uncensored-IQ4_XS.gguf Q6_K

1. 1.Split into <50 GB shards with llama-gguf-split.

License

Inherited from the base model MiniMaxAI/MiniMax-M3.