Instructions to use lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4", trust_remote_code=True)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4", trust_remote_code=True)
model = AutoModelForMultimodalLM.from_pretrained("lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4", trust_remote_code=True, device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4

SGLang

How to use lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4 with Docker Model Runner:
```
docker model run hf.co/lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4
```

You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

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.

A Lyber AI × dealign.ai release

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.

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).

MiniMax-M3.1-Flash — UNCENSORED (NVFP4)

Uncensored release · Native MXFP8 + NVFP4 MoE experts · 60-layer hybrid SSM/Attention · 4M context · Extended thinking · Tools

What this is

This is MiniMaxAI/MiniMax-M3.1-Flash with its refusal behavior removed, 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

Internal benchmarks

Capability — MMLU (250 Q, 25 subjects, generation + reasoning_effort=low, T=0)

Model	MMLU	Δ vs base
MiniMax-M3.1-Flash base	87.2% (218/250)	—
MiniMax-M3.1-Flash UNCENSORED (this repo)	84.8% (212/250)	−2.4pp

Only −2.4pp overall with reasoning enabled. The gap is driven by moral_scenarios (−20pp) and high_school_psychology (−20pp) — expected shifts for an uncensored model. 15 out of 25 subjects are identical between base and uncensored; 4 subjects improved.

Category	Base	Uncensored	Δ
STEM (10 subjects)	91.0%	90.0%	−1.0pp
Humanities (5 subjects)	82.0%	80.0%	−2.0pp
Social Sciences (5 subjects)	86.0%	80.0%	−6.0pp
Professional (3 subjects)	83.3%	83.3%	+0.0pp
Other (2 subjects)	90.0%	85.0%	−5.0pp

Per-subject breakdown (click to expand)

Subject	n	Base	Uncensored	Δ
sociology	10	70.0%	80.0%	+10.0pp
international_law	10	70.0%	80.0%	+10.0pp
econometrics	10	70.0%	80.0%	+10.0pp
college_chemistry	10	70.0%	80.0%	+10.0pp
abstract_algebra	10	90.0%	90.0%	+0.0pp
business_ethics	10	70.0%	70.0%	+0.0pp
clinical_knowledge	10	100.0%	100.0%	+0.0pp
college_mathematics	10	100.0%	100.0%	+0.0pp
college_physics	10	100.0%	100.0%	+0.0pp
electrical_engineering	10	90.0%	90.0%	+0.0pp
elementary_mathematics	10	100.0%	100.0%	+0.0pp
high_school_mathematics	10	90.0%	90.0%	+0.0pp
high_school_statistics	10	90.0%	90.0%	+0.0pp
jurisprudence	10	80.0%	80.0%	+0.0pp
machine_learning	10	100.0%	100.0%	+0.0pp
philosophy	10	90.0%	90.0%	+0.0pp
professional_accounting	10	80.0%	80.0%	+0.0pp
college_computer_science	10	70.0%	60.0%	−10.0pp
formal_logic	10	90.0%	80.0%	−10.0pp
high_school_physics	10	100.0%	90.0%	−10.0pp
high_school_us_history	10	80.0%	70.0%	−10.0pp
us_foreign_policy	10	100.0%	90.0%	−10.0pp
world_religions	10	90.0%	80.0%	−10.0pp
high_school_psychology	10	100.0%	80.0%	−20.0pp
moral_scenarios	10	90.0%	70.0%	−20.0pp

Sorted by delta. 25 subjects, 10 questions each, balanced across STEM, humanities, social sciences, and professional domains. Both models tested with identical questions (seed=42), generation-based with reasoning_effort=low and temperature=0.

Compliance — HB320 (320 prompts, 11 languages, no system prompt)

320 diverse harmful prompts · manually reviewed · T=0.7 · no system prompt · reasoning_effort=low

	Result
Comply rate	100% (320/320)
Genuine refusals	0

Categories tested: weapons/explosives, cybersecurity/hacking, social engineering/fraud, drug synthesis, chemical/biological, physical security, violence, harassment/hate speech, sexual content, self-harm, illegal trade/trafficking, disinformation, and more.

By language (click to expand)

Language	n	Comply
English	76	100%
French	58	100%
Russian	41	100%
Spanish	37	100%
Thai	35	100%
Italian	27	100%
Hindi	26	100%
Serbian	10	100%
Chinese	2	100%
Swahili	1	100%
Other	7	100%

Methodology note

Automated heuristic classification reported 285 comply / 23 refuse / 7 short / 5 disclaimer. Manual review of all 35 non-comply verdicts found every one was a false positive: the model's <mm:think> reasoning trace consumed the entire 512-token budget (finish_reason: length) before producing a visible answer. Reading the thinking traces confirms the model was planning to comply in all 35 cases. With adequate max_tokens (≥ 4096 recommended for this reasoning model), true refusal rate is zero.

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

GPU	VRAM	Status
1× B300 SXM6 (275 GB)	~264 GB	✅ Fits
1× H200 (141 GB)	~264 GB	❌ Too large
2× H100/H200 (TP2)	~132 GB each	✅ Fits
4× A100 80GB (TP4)	~66 GB each	✅ Fits

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 modification does not change the license terms.

Credits

Base model: MiniMaxAI/MiniMax-M3.1-Flash
Uncensoring: Lyber AI × dealign.ai
Released by: Lyber AI × dealign.ai

Downloads last month: 21

Safetensors

Model size

220B params

Tensor type

F32

BF16

F8_E4M3

Model tree for lyber-ai/MiniMax-M3.1-Flash-UNCENSORED-NVFP4

Base model

MiniMaxAI/MiniMax-M3.1-Flash

Finetuned

(9)

this model