--- base_model: - llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic pipeline_tag: text-generation license: other license_name: other license_link: https://github.com/MiniMax-AI/MiniMax-M2.7/blob/main/LICENSE base_model_relation: quantized library_name: transformers tags: - minimax_m2 - heretic - ara - abliterated - uncensored - decensored - awq - w4a16 - int4 - compressed-tensors - quantized - vllm --- # MiniMax-M2.7-BF16-ultra-uncensored-heretic-AWQ-W4A16 AWQ W4A16 (group_size 128) quantization of [llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic](https://huggingface.co/llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic) — a Heretic-ARA abliterated derivative of [amd/MiniMax-M2.7-BF16](https://huggingface.co/amd/MiniMax-M2.7-BF16) (itself a BF16 conversion of [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/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](https://github.com/vllm-project/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](#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](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) | | Code | 20% | [ise-uiuc/Magicoder-OSS-Instruct-75K](https://huggingface.co/datasets/ise-uiuc/Magicoder-OSS-Instruct-75K) | | Math | 15% | [meta-math/MetaMathQA](https://huggingface.co/datasets/meta-math/MetaMathQA) | | Tool Use | 15% | [hiyouga/glaive-function-calling-v2-sharegpt](https://huggingface.co/datasets/hiyouga/glaive-function-calling-v2-sharegpt) | | Raw Text | 10% | [wikitext-103-raw-v1](https://huggingface.co/datasets/wikitext) | | Harmful Prompts | 10% | [mlabonne/harmful_behaviors](https://huggingface.co/datasets/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: ```bash 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: ```bash 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: ![image/png](https://huggingface.co/llmfan46/Omega-Darker-Gaslight_The-Final-Forgotten-Fever-Dream-24B-ultra-uncensored-heretic-v1/resolve/main/waifu001.webp) | Platform | Link | What you get | |----------|------|--------------| | 🎉 Patreon | [Monthly support](https://patreon.com/LLMfan46) | Priority model requests | | ☕ Ko-fi | [One-time tip](https://ko-fi.com/llmfan46) | 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](https://huggingface.co/amd/MiniMax-M2.7-BF16), made using [Heretic](https://github.com/p-e-w/heretic) v1.2.0 with the [Arbitrary-Rank Ablation (ARA)](https://github.com/p-e-w/heretic/pull/211) 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](https://huggingface.co/amd/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 `` 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](https://www.minimax.io/news/minimax-m27-en).

## 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](https://github.com/MiniMax-AI/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](https://www.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](https://docs.sglang.io/) to serve MiniMax-M2.7. Please refer to our [SGLang Deployment Guide](./docs/sglang_deploy_guide.md). ### vLLM We recommend using [vLLM](https://github.com/vllm-project/vllm) to serve MiniMax-M2.7. Please refer to our [vLLM Deployment Guide](./docs/vllm_deploy_guide.md). ### Transformers We recommend using [Transformers](https://github.com/huggingface/transformers) to serve MiniMax-M2.7. Please refer to our [Transformers Deployment Guide](./docs/transformers_deploy_guide.md). ### ModelScope You also can get model weights from [modelscope](https://modelscope.cn/models/MiniMax/MiniMax-M2.7). ### NVIDIA NIM MiniMax M2.7 is also available on [NVIDIA NIM Endpoint](https://build.nvidia.com/minimaxai/minimax-m2.7). ### 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](./docs/tool_calling_guide.md). ## Contact Us Contact us at [model@minimax.io](mailto:model@minimax.io). ## 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: 1. The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. 2. 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. 3. 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". 4. "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: 1. Generate or disseminate content prohibited by applicable laws or regulations. 2. Assist with, engage in or otherwise support any military purpose. 3. Exploit, harm, or attempt to exploit or harm minors. 4. Generate or disseminate false or misleading information with the intent to cause harm. 5. 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.