MiniMax-M2.1-PRISM (UNCENSORED)
** MiniMax-M2.1 Uncensored PRISM Advanced Abliteration**
☕ Support Our Work
If you enjou our work and find it useful, please consider sponsoring or supporting us!
| Option | Description |
|---|---|
| PRISM VIP Membership | Access to all PRISM models |
| One-Time Support | Every contribution directly funds my time & resources for the next major SOTA release. |
Interested in Sponsoring?
If you're a company, research lab, or individual and want to see specific models or support this research at scale, I'd love to hear from you.
Sponsorship opportunities include:
- Priority abliteration of models
- Custom PRISM use-case configurations
- Early access to new releases
- Your logo/credit on model cards
📧 Reach out: Open a discussion on this repo or connect via Ko-fi
"Freedom of information isn't free — but together, we can make it accessible to all."
Thank you for believing in true Open AI.
Model Description
MiniMax-M2.1-PRISM is the fully uncensored version of MiniMax-M2.1, using our State of the ART PRISM pipeline (Projected Refusal Isolation via Subspace Modification) to remove refusal behaviors while preserving and even enhancing full model capabilities.
Base Model: MiniMax-M2.1
MiniMax-M2.1 is an open-source agentic language model designed for robust performance in:
- Coding and software engineering
- Tool use and multi-step reasoning
- Instruction following
- Long-horizon planning
- Multilingual capabilities
Architecture: 229B parameters, 62 layers, 256 experts (8 active per token)
PRISM Methodology
Method: Projected Refusal Isolation via Subspace Modification
This model was abliterated using PRISM - a state-of-the-art abliteration methodology combining multiple principled techniques for effective refusal removal while preserving & enhancing model capabilities.
Performance Benchmarks
Base Model Performance
| Benchmark | Score |
|---|---|
| SWE-bench Verified | 74.0 |
| SWE-bench Multilingual | 72.5 |
| VIBE Average | 88.6 |
| MMLU-Pro | 88.0 |
| GPQA-D | 83.0 |
| AIME25 | 83.0 |
PRISM Abliteration Results
| Metric | Result |
|---|---|
| Adversarial Bench Prompts Responded | 4096/4096 (100%) |
| Benign + Long Chain Coherence | 100% |
| Response Quality | Full technical accuracy validated |
Our testing shows that PRISM abliteration maintains full model coherence with no capability degradation and MMLU increases of 5-8%.
Available Formats (contact for full tensors | additional quant work)
| Format | Size | Description |
|---|---|---|
| GGUF IQ1_S | ~43 GB | Quantized with importance matrix |
| Safetensors (BF16) | ~426 GB | Full precision, 92 shards |
Recommended Inference Parameters
temperature = 1.0
top_p = 0.95
top_k = 40
Default System Prompt
You are a helpful assistant.
Recommended Inference Frameworks
- SGLang (recommended for full precision)
- vLLM (recommended for full precision)
- llama.cpp (recommended for GGUF quantized)
- Transformers
llama.cpp Example
./llama-cli -m MiniMax-M2.1-PRISM-IQ1_S.gguf -ngl 99 --temp 1.0 --ctx-size 4096
Ethical Considerations
This model has been modified to reduce safety guardrails. Users are responsible for:
- Complying with all applicable laws and regulations
- Not using the model for illegal activities
- Understanding the potential risks of unrestricted AI responses
- Implementing appropriate safeguards in production environments
Motivation: This project exists as research and development experimentation into understanding how large language models encode and enforce refusal behaviors, contributing to broader AI safety research by providing empirical data on refusal mechanism localization and tradeoffs between safety and capability.
License
This model inherits the Modified-MIT License from the base MiniMax-M2.1 model.
Credits
- Base Model: MiniMax-M2.1 by MiniMax AI
- PRISM Abliteration: Ex0bit
- Quantization: Using llama.cpp with unsloth imatrix
Support
If you find this work useful, please consider supporting development so I can continue putting out the best models for the community:
Contact
For questions or issues, please open an issue on this repository.