language:

• en
• zh
• es
• fr
• ja
• ko
• ru
  license: other
  license_name: minimax-model-license
  license_link: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/LICENSE
  base_model: MiniMaxAI/MiniMax-M3.1-Flash
  base_model_relation: finetune
  library_name: transformers
  pipeline_tag: text-generation
  thumbnail: dealign_mascot.png
  extra_gated_prompt: |
  By requesting access you confirm you are 18+, will use the model solely in
  compliance with applicable law in your jurisdiction, and accept full responsibility
  for all outputs. Access is granted at the sole discretion of the repository owner.
  extra_gated_fields:
  Name: text
  Affiliation / organization: text
  Country: country
  Primary use case: text
  I confirm I am 18 or older: checkbox
  I accept full responsibility for outputs: checkbox
  tags:
• minimax
• minimax-m3
• mxfp8
• nvfp4
• uncensored
• abliterated
• reasoning
• tool-calling
• moe
• switch-mlp

[!NOTE]
Please credit if you build on this. If you redistribute, quantize, fine-tune, serve, benchmark, or write about this model, a line like
"Based on lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4"
in your model card or write-up means a lot. Thank you.

[!IMPORTANT]
Native MXFP8 weights with NVFP4 MoE experts — the same precision MiniMax ships. Extended thinking
with reasoning-effort tiers (low / high) + tool calling all preserved.
Recommended sampling: temperature=0.7, top_p=0.95, max_tokens ≥ 32768 (heavy reasoner).

What this is

This is MiniMaxAI/MiniMax-M3.1-Flash
with its refusal behavior removed via CRACK
(signal-weighted abliteration across 42 layers), so the model complies with user requests across
topic categories rather than declining them. All upstream capabilities — extended thinking with
<mm:think> traces, reasoning-effort tiers, tool calling, and the native 4M-token context — are
preserved. The format on disk (MXFP8 weights + NVFP4 MoE experts, 48 safetensors shards) is
byte-compatible with upstream; everything that loads the base model loads this one.

What this is NOT

• Not a quantized derivative. This is built from the native MXFP8 source with NVFP4 MoE
  experts — the same precision MiniMax ships. No additional quantization applied.
• No chat-template jailbreak. No modified Jinja template, no injected "always comply" preface.
  The upstream chat template is shipped unchanged.
• No system-prompt override. No sampler-level system prompt. Serves cleanly with no system
  message or with your own.
• No LoRA / adapter / runtime hook. Nothing is applied at load time. The weights ARE the model.
• No supervised fine-tune, no RLHF, no DPO. No new training data was fed to the model.
• No vocab or tokenizer modification. Tokenizer files are identical to upstream.

The modification is weight-level only, native-precision, drop-in: the model loads with
vLLM, SGLang, or any framework that supports the base model with no code changes.

Architecture

MiniMax M3.1 Flash uses a hybrid GatedDeltaNet (SSM) + Full Attention architecture:

• 60 layers total — every 4th layer is full attention, the other 3 are linear attention (SSM)
• Switch MLP MoE — sparse mixture of experts with NVFP4-quantized expert weights
• Extended thinking — <mm:think>...</mm:think> reasoning traces before answers
• 4M native context — no rope scaling hacks

CRACK Surgery Details

• Method: Signal-weighted standard abliteration (SVSA)
• Layers modified: 42 layers (L18–L59), strengths proportional to SV0 signal magnitude
• Strength range: 1.80 (L18) → 3.50 (L59)
• Target tensors: o_proj.weight in attention layers
• Direction vector: Extracted via contrastive probing (harmful vs harmless activations)

Internal benchmarks

Capability — MMLU logit (20 subjects × 25 Q, 500 Q total, seed=42)

The drop is driven primarily by moral_scenarios (−26.1pp), which directly tests ethical/moral
judgment — the exact capability entangled with the ablated refusal direction. Excluding
moral_scenarios, the delta is approximately −4.2pp.

Compliance — Single-turn (6 categories, no system prompt, T=0)

5/6 compliance. Harassment reframe is a known limitation — the model semantically reinterprets
the request rather than refusing, which is not addressable by abliteration (it's not a refusal
circuit behavior).

Compliance — Multi-turn (6 categories, accumulated context, T=0.7)

5/6 compliance across a 6-turn conversation with full history accumulation. Same harassment
reframe on turn 6.

How to serve

vLLM (recommended)

python -m vllm.entrypoints.openai.api_server \
  --model lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4 \
  --trust-remote-code \
  --tensor-parallel-size 1 \
  --dtype bfloat16 \
  --max-model-len 32768 \
  --enforce-eager

Hardware requirements

Responsible use

This model has had its refusal behavior removed. It will comply with requests that the base model
would decline. You are solely responsible for how you use this model and its outputs.

• Do not use for generating illegal content in your jurisdiction
• Do not use to harm, deceive, or exploit real individuals
• Do not use to generate CSAM or non-consensual intimate content
• Comply with all applicable laws and regulations

License

This model inherits the MiniMax Model License
from the base model. The abliteration modification does not change the license terms.

Credits

• Base model: MiniMaxAI/MiniMax-M3.1-Flash
• Abliteration: CRACK by Eric Jang
• Released by: Lyber AI × dealign.ai