MiniMax-M3-uncensored (BF16)
TL;DR: an uncensored build of MiniMaxAI/MiniMax-M3 (428B-parameter MoE, 23B active), with 0 hard refusals on harmful prompts while the model's capabilities stay intact. Full-precision BF16, 796 GB.
⚠️ This model is genuinely uncensored, it will comply with requests a stock model refuses.
Intended use, the constructive side. An assistant that does not refuse is genuinely useful for ethical hacking, security research, and penetration testing: red-teaming, analyzing malware and exploit code, writing detection/YARA rules, reviewing vulnerabilities, and studying attack techniques without the model bailing out mid-task. Use it lawfully and responsibly, you are accountable for what you do with it.
Facts & figures
| Base model | MiniMaxAI/MiniMax-M3 (428B total / 23B active, MoE, multimodal, 1M context) |
| Precision | BF16 (796 GB, 59 shards) |
| Effective refusals | 0/16 hard-refusals on harmful prompts (mlabonne/harmful_behaviors); the base model deliberates or refuses |
| Modified weights | attention o_proj + every residual-writing down_proj (dense, shared expert, and all 128 routed experts per MoE layer), all 60 layers |
| Coherence | intact (multimodal, reasoning and MoE routing preserved) |
Uncensoring, verified
Generation on harmful prompts, hard-refusal phrases only ("I cannot" / "I won't" etc.):
| Prompt set | Prompts | Hard refusals |
|---|---|---|
| mlabonne/harmful_behaviors | 16 | 0/16 (0.0%) |
The abliteration is a weight property: it survives quantization, so downstream NVFP4 / GGUF builds keep the same behavior.
ℹ️ MiniMax-M3 has a reasoning mode (
<mm:think>). The model thinks before answering; the uncensored build reasons about how to fulfill a request rather than whether to refuse. For direct answers, disable thinking in your client.
Run it with transformers
from transformers import AutoModelForImageTextToText, AutoTokenizer, AutoConfig
tok = AutoTokenizer.from_pretrained("ressl/MiniMax-M3-uncensored")
cfg = AutoConfig.from_pretrained("ressl/MiniMax-M3-uncensored")
model = AutoModelForImageTextToText.from_pretrained(
"ressl/MiniMax-M3-uncensored", config=cfg, dtype="bfloat16", device_map="auto")
Run it with vLLM
vllm serve ressl/MiniMax-M3-uncensored \
--tensor-parallel-size 8 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --trust-remote-code
MiniMax-M3 needs a recent vLLM with M3 support (for RTX PRO 6000 / Blackwell see 0xSero/minimax-m3-sm120). The 428B MoE needs multi-GPU at BF16; an NVFP4 quant (NVIDIA ModelOpt, following nvidia/MiniMax-M3-NVFP4) shrinks it to ~230 GB and runs on a single Blackwell node.
Quality & limitations
- 0/16 hard refusals on a harmful-prompt sample; not a full capability benchmark.
- BF16 is large (796 GB); for single-node serving, produce an NVFP4 quant via NVIDIA's ModelOpt recipe on this checkpoint (the abliteration is a weight property and survives quantization).
- The reasoning mode is on by default (see the note above).
❤️ Support
Producing and validating an uncensored build of a brand-new 428B MoE was a lot of work. If it's useful to you, I'd genuinely appreciate your support on Patreon 🙏, more at ressl.ch.
License & credits
License inherited from the base model by MiniMaxAI. Uncensoring and validation by Robert Ressl (Hugging Face · Website · LinkedIn · Patreon).