Support & Community

---

MiniMax-M3 — Abliterated (BF16)

Overview

This is an abliterated (uncensored) build of MiniMaxAI/MiniMax-M3 — the full bfloat16 weights with the model's refusal behavior removed, while keeping its reasoning, multilingual, coding, and multimodal abilities intact. Legitimate safety and security-analysis engagement is preserved, as is tool use — the model simply stops reflexively refusing.

MiniMax-M3 is a large multimodal Mixture-of-Experts model with a vision tower and a built-in reasoning ("thinking") mode. The architecture, tokenizer, and chat template are unchanged, so this is a drop-in replacement for the base model.

Usage

Serve with vLLM (a MiniMax-M3-capable build is required):

MODEL=OpenYourMind/Minimax-M3-abliterated-clean

vllm serve "$MODEL" \
  --tensor-parallel-size 8 \
  --block-size 128 \
  --reasoning-parser minimax_m3 \
  --tool-call-parser minimax_m3 \
  --enable-auto-tool-choice

The model wraps its reasoning in <mm:think> … </mm:think>; use the minimax_m3 reasoning parser to surface it. --block-size 128 is required for MiniMax Sparse Attention.

Files

Total on disk: ~854 GB (bfloat16).

Hardware

These are full BF16 weights — plan for a multi-GPU / multi-node deployment (e.g. 8×/16× 80 GB-class accelerators), or quantize (NVFP4 / MXFP4 / FP8) to fit smaller setups. Quantized builds may follow.

Notes

• License: Other (inherits from the MiniMax-M3 base license)
• Base model: MiniMaxAI/MiniMax-M3
• Modality: text + vision (image-text-to-text) with reasoning / thinking mode

Disclaimer

Use is the responsibility of the user. Ensure your usage complies with applicable laws, platform rules, and deployment requirements.