base_model: ressl/MiniMax-M3-uncensored
license: other
tags:

• gguf
• minimax-m3
• uncensored
• abliterated
• imatrix

MiniMax-M3-uncensored-GGUF

Static GGUF quants of ressl/MiniMax-M3-uncensored,
the abliterated (refusal-removed) build of MiniMaxAI/MiniMax-M3
(427B total / 10B active MoE, 128 experts, 4 active). Unpruned, text-only.

[!NOTE]
All credit for the model goes to ressl (abliteration of o_proj and every
down_proj, all 60 layers, routers untouched) and MiniMaxAI (base model).
The importance matrix and the per-tensor quant recipe are AesSedai's, taken
verbatim from AesSedai/MiniMax-M3-GGUF so these files are
quant-matched tensor-for-tensor to their stock MiniMax-M3 quants. Thank you.

What was done

1. ressl/MiniMax-M3-uncensored bf16 safetensors → Q8_0 GGUF with convert_hf_to_gguf.py (llama.cpp b10819).

2. Each smaller size cut from that Q8_0 with llama-quantize --allow-requantize using AesSedai's
   imatrix.gguf (bartowski calibration_datav5, 49 chunks) and their mixture: the ftype sets the dense
   tensors, and --tensor-type overrides push only the routed experts down. Example for IQ4_XS:

   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
   

3. Files are split into <50 GB shards with llama-gguf-split; point llama.cpp at shard 00001 and it loads the set.

Quants

Uploading in this order; the table fills in as each size lands. PPL and KLD are measured against this
model's own Q8_0 on wiki.test.raw (512-token chunks), so they read as quantization loss of the
uncensored weights, not as a comparison to the stock model.

Serving

Thinking is on by default in MiniMax-M3's template; pass chat_template_kwargs: {"thinking_mode": "disabled"}
for the non-thinking mode. The IQ4_XS fits a 256 GB Apple Silicon box.

Caliper

This build is measured on the Caliper creative-writing leaderboard
against the stock MiniMax-M3 served from the same AesSedai recipe.