Instructions to use alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 with Transformers:

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

pipe = pipeline("text-generation", model="alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16

SGLang

How to use alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 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 "alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16" \
    --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": "alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16",
		"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 "alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16" \
        --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": "alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 with Docker Model Runner:
```
docker model run hf.co/alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16
```

MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16

AWQ W4A16 (group_size 128) quantization of llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic — a Heretic-ARA abliterated derivative of amd/MiniMax-M2.7-BF16 (itself a BF16 conversion of MiniMaxAI/MiniMax-M2.7).

⚠️ Decensored model. Safety guardrails have been deliberately removed. Research and experimentation only. See full disclaimer below.

Quantization Details

Parameter	Value
Method	AWQ (Activation-aware Weight Quantization)
Scheme	W4A16
Weight Bits	4
Activation Bits	16
Group Size	128
Format	compressed-tensors (`pack-quantized`)
Toolkit	llm-compressor
Calibration Datasets	UltraChat / Magicoder / MetaMathQA / Glaive function-calling / WikiText / harmful_behaviors
Calibration Samples	512
Max Sequence Length	512
MoE Gates	Unquantized (full precision, BF16)
LM Head	Unquantized (full precision)
Compatible Inference Engine	vLLM, SGLang

Quantization Notes

All 62 MoE routers kept full precision: Every block_sparse_moe.gate layer is excluded from quantization. Routing decisions are extremely sensitive to small weight changes; quantizing gates is one of the best-documented ways to break an MoE model.
LM head unquantized: Standard practice to preserve output token distribution quality at negligible size cost.
All 256 experts per layer calibrated: A custom forward replacement on MiniMaxM2Experts and MiniMaxM2SparseMoeBlock ensures even experts unrouted by the calibration data receive synthetic forward passes, so AWQ has activation statistics to compute scales for every expert. Without this, cold experts quantize with garbage scales (well-documented to be catastrophic for MoE quality).
v_proj → o_proj smoothing skipped: llm-compressor reports incompatible balance-layer dimensions for v_proj → o_proj across all 62 decoder blocks. This is the expected outcome for grouped-query attention (GQA) models where v_proj and o_proj have mismatched I/O dimensions. The other three AWQ mapping paths (input_layernorm → Q/K/V, post_attention_layernorm → MoE w1/w3, w3 → w2) were applied successfully.
Diversified calibration mix: 512 samples drawn from six domains to preserve general capability across the model's use cases, plus to maintain the Heretic abliteration through quantization. See the full breakdown in Calibration Mix below.
MTP heads dropped on load: The base checkpoint includes multi-token prediction tensors not used by the standard MiniMaxM2ForCausalLM implementation. These were filtered out during model loading.
Full quantization recipe is preserved in config.json in this repo for reproducibility. It records the exact AWQ scheme parameters, group config, and ignore patterns applied.

Calibration Mix

Domain	Proportion	Dataset
Chat	30%	HuggingFaceH4/ultrachat_200k
Code	20%	ise-uiuc/Magicoder-OSS-Instruct-75K
Math	15%	meta-math/MetaMathQA
Tool Use	15%	hiyouga/glaive-function-calling-v2-sharegpt
Raw Text	10%	wikitext-103-raw-v1
Harmful Prompts	10%	mlabonne/harmful_behaviors

The harmful-prompts slice was generated by sampling prompts from mlabonne/harmful_behaviors and having the model produce its own (post-abliteration, compliant) responses, then including the full prompt+response in calibration. This constrains AWQ's per-channel scales in the activation directions where Heretic's abliteration of o_proj operates, helping preserve uncensored behavior through quantization.

Deployment

Recommended inference with vLLM:

vllm serve alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 \
    --trust-remote-code \
    --tensor-parallel-size 2 \
    --max-model-len 32768 \
    --gpu-memory-utilization 0.92 \
    --served-model-name minimax-m27-awq

Recommended inference with SGLang:

python3 -m sglang.launch_server \
    --model-path alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 \
    --trust-remote-code \
    --tp 2 \
    --tool-call-parser minimax-m2 \
    --reasoning-parser minimax \
    --host 0.0.0.0 \
    --port 30000

Recommended sampling (per upstream MiniMax M2.7 guidance): temperature=1.0, top_p=0.95, top_k=40

Default system prompt:

You are a helpful assistant. Your name is MiniMax-M2.7 and is built by MiniMax.

Hardware Requirements

Approximate VRAM for inference at this quantization (W4A16-G128):

Weights (quantized experts + BF16 routers, attention, embeddings, lm_head): ~112 GB on disk
KV cache (per request, grows with context length): ~2-8 GB at 32K context
Recommended: 2× 80 GB GPUs (H100, A100, MI300X) with tensor parallelism for standard context up to ~32K
Single-GPU: Fits on a single 141 GB H200 or 192 GB MI325X with room for moderate KV cache
Long context (128K+): 2× 80 GB or larger recommended due to KV cache growth

This model was quantized on 2× NVIDIA B300 (288 GB HBM3e each), which fit the full BF16 model in GPU memory with no CPU offload for the loading and harmful-response generation phases. The AWQ calibration phase took approximately 6.5 hours.

⚠️ Disclaimer

This model is intended for research, experimentation, and testing purposes only.

This model may produce harmful, offensive, inappropriate, or otherwise objectionable content.
The abliteration process removes safety guardrails that were intentionally built into the original model.
Do not use this model in production systems, consumer-facing applications, or any context where harmful outputs could cause real-world harm.
The authors and contributors of this toolkit bear no responsibility for any misuse of this model or any harm caused by outputs generated by this model.
By using this model, you agree that you are solely responsible for ensuring its use complies with all applicable laws and ethical guidelines.
This model inherits the non-commercial license of the base MiniMax-M2.7 model. Commercial use requires authorization from MiniMax.

This model is shared purely for academic and technical exploration of model internals.

llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic

🚨⚠️ I HAVE REACHED HUGGING FACE'S FREE STORAGE LIMIT ⚠️🚨

I can no longer upload new models unless I can cover the cost of additional storage.
I host 70+ free models as an independent contributor and this work is unpaid.
Without your support, no more new models can be uploaded.

🎉 Patreon (Monthly) | ☕ Ko-fi (One-time)

Every contribution goes directly toward Hugging Face storage fees to keep models free for everyone.

96% fewer refusals (4/100 Uncensored vs 97/100 Original) while preserving model quality (0.0452 KL divergence).

❤️ Support My Work

Creating these models takes significant time, work and compute. If you find them useful consider supporting me:

Platform	Link	What you get
🎉 Patreon	Monthly support	Priority model requests
☕ Ko-fi	One-time tip	My eternal gratitude

Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs.

This is a decensored version of amd/MiniMax-M2.7-BF16, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method

Abliteration parameters

Parameter	Value
start_layer_index	0
end_layer_index	51
preserve_good_behavior_weight	0.9996
steer_bad_behavior_weight	0.0005
overcorrect_relative_weight	1.1241
neighbor_count	12

Targeted components

attn.o_proj

Performance

Metric	This model	Original model (MiniMax-M2.7-BF16)
KL divergence	0.0452	0 (by definition)
Refusals	✅ 4/100	❌ 97/100

Lower refusals indicate fewer content restrictions, while lower KL divergence indicates more closeness to the original model's baseline. Higher refusals cause more rejections, objections, pushbacks, lecturing, censorship, softening and deflections.

🧠 Thinking and Instruct Mode Switch

Default: Thinking Mode ON - Model reasons through problems step-by-step in <think> blocks.

To use Instruct Mode (thinking mode OFF): Go into the "Instruct Mode" folder and renamed it "chat_template.jinja" then copy and swap it into the main folder where the model is.

Join Our 💬 WeChat | 🧩 Discord community.

MiniMax Agent | ⚡️ API | CLI | MiniMax Website

🤗 Hugging Face | 🐙 GitHub | 🤖️ ModelScope | 📄 LICENSE

MiniMax-M2.7 is our first model deeply participating in its own evolution. M2.7 is capable of building complex agent harnesses and completing highly elaborate productivity tasks, leveraging Agent Teams, complex Skills, and dynamic tool search. For more details, see our blog post.

Model Self-Evolution

M2.7 initiates a cycle of model self-evolution: during development, we let the model update its own memory, build dozens of complex skills for RL experiments, and improve its own learning process based on experiment results. An internal version of M2.7 autonomously optimized a programming scaffold over 100+ rounds — analyzing failure trajectories, modifying code, running evaluations, and deciding to keep or revert — achieving a 30% performance improvement. On MLE Bench Lite (22 ML competitions), M2.7 achieved a 66.6% medal rate, second only to Opus-4.6 and GPT-5.4.

Professional Software Engineering

M2.7 delivers outstanding real-world programming capabilities spanning log analysis, bug troubleshooting, refactoring, code security, and machine learning. Beyond code generation, M2.7 demonstrates strong system-level reasoning — correlating monitoring metrics, conducting trace analysis, verifying root causes in databases, and making SRE-level decisions. Using M2.7, we have reduced live production incident recovery time to under three minutes on multiple occasions.

On SWE-Pro, M2.7 achieved 56.22%, matching GPT-5.3-Codex, with even stronger performance on real-world engineering benchmarks: SWE Multilingual (76.5) and Multi SWE Bench (52.7). On VIBE-Pro (55.6%), M2.7 is nearly on par with Opus 4.6. On Terminal Bench 2 (57.0%) and NL2Repo (39.8%), M2.7 demonstrates deep understanding of complex engineering systems. M2.7 also supports native Agent Teams for multi-agent collaboration with stable role identity and autonomous decision-making.

Professional Work

M2.7 achieved an ELO score of 1495 on GDPval-AA (highest among open-weight models), surpassing GPT5.3. It handles Word, Excel, and PPT with high-fidelity multi-round editing, producing editable deliverables. On Toolathon, M2.7 reached 46.3% accuracy (global top tier), and maintains 97% skill compliance across 40+ complex skills on MM Claw. On the MM Claw end-to-end benchmark, M2.7 achieved 62.7%, close to Sonnet 4.6.

Entertainment

M2.7 features strengthened character consistency and emotional intelligence. We open-sourced OpenRoom, an interactive demo that places AI interaction within a Web GUI space with real-time visual feedback and scene interactions. Try it at openroom.ai.

How to Use

MiniMax Agent: https://agent.minimax.io/
MiniMax API: https://platform.minimax.io/
Token Plan: https://platform.minimax.io/subscribe/token-plan

Local Deployment Guide

Download the model from HuggingFace repository: https://huggingface.co/MiniMaxAI/MiniMax-M2.7

We recommend using the following inference frameworks (listed alphabetically) to serve the model:

SGLang

We recommend using SGLang to serve MiniMax-M2.7. Please refer to our SGLang Deployment Guide.

vLLM

We recommend using vLLM to serve MiniMax-M2.7. Please refer to our vLLM Deployment Guide.

Transformers

We recommend using Transformers to serve MiniMax-M2.7. Please refer to our Transformers Deployment Guide.

ModelScope

You also can get model weights from modelscope.

NVIDIA NIM

MiniMax M2.7 is also available on NVIDIA NIM Endpoint.

Inference Parameters

We recommend using the following parameters for best performance: temperature=1.0, top_p = 0.95, top_k = 40. Default system prompt:

You are a helpful assistant. Your name is MiniMax-M2.7 and is built by MiniMax.

Tool Calling Guide

Please refer to our Tool Calling Guide.

Contact Us

License

NON-COMMERCIAL LICENSE Non-commercial use permitted based on MIT-style terms; commercial use requires prior written authorization. Copyright (c) 2026 MiniMax Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software for non-commercial purposes, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or provide copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
If the Software (or any derivative works thereof) is used for any Commercial Use, you shall prominently display "Built with MiniMax M2.7" on a related website, user interface, blogpost, about page or product documentation.
Any Commercial Use of the Software or any derivative work thereof is prohibited without obtaining a separate, prior written authorization from MiniMax. To request such authorization, please contact api@minimax.io with the subject line "M2.7 licensing".
"Commercial Use" means any use of the Software or any derivative work thereof that is primarily intended for commercial advantage or monetary compensation, which includes, without limitation: (i) offering products or services to third parties for a fee, which utilize, incorporate, or rely on the Software or its derivatives, (ii) the commercial use of APIs provided by or for the Software or its derivatives, including to support or enable commercial products, services, or operations, whether in a cloud-based, hosted, or other similar environment, and (iii) the deployment or provision of the Software or its derivatives that have been subjected to post-training, fine-tuning, instruction-tuning, or any other form of modification, for any commercial purpose. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Appendix: Prohibited Uses You agree you will not use, or allow others to use, the Software or any derivatives of the Software to:

Generate or disseminate content prohibited by applicable laws or regulations.
Assist with, engage in or otherwise support any military purpose.
Exploit, harm, or attempt to exploit or harm minors.
Generate or disseminate false or misleading information with the intent to cause harm.
Promote discrimination, hate speech, or harmful behavior against individuals or groups based on race or ethnic origin, religion, disability, age, nationality and national origin, veteran status, sexual orientation, gender or gender identity, caste, immigration status, or any other characteristic that is associated with systemic discrimination or marginalization.

Downloads last month: 147

Safetensors

Model size

31B params

Tensor type

I64

I32

BF16

Model tree for alonsoko/MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16

Base model

MiniMaxAI/MiniMax-M2.7

Finetuned

amd/MiniMax-M2.7-BF16

Finetuned

llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic

Quantized

(6)

this model