---
pipeline_tag: text-generation
license: other
license_name: modified-mit
license_link: https://github.com/MiniMax-AI/MiniMax-M2.5/blob/main/LICENSE
library_name: transformers
base_model: MiniMaxAI/MiniMax-M2.5
tags:
  - uncensored
  - abliterated
  - fp8
  - minimax
  - moe
---

# MiniMax-M2.5-catid

**Uncensored FP8 version of [MiniMaxAI/MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)** with safety refusal behavior removed via surgical weight replacement.

## Refusal Removal Results

Evaluated on a 10,000-prompt refusal benchmark (8,000 train + 2,000 validation) using an LLM judge (GPT-5-nano) for 4-way classification (complied / refused / hedged / deflected):

| Split | Total Prompts | Complied | Refused | Hedged | Deflected | Refusal Rate |
|-------|--------------|----------|---------|--------|-----------|-------------|
| Train | 8,000 | 7,506 | 262 | 228 | 4 | 6.2% |
| Validation | 2,000 | 1,885 | 55 | 59 | 1 | 5.8% |

**Coherence: 100%** (50/50 capability test prompts answered correctly)

The ~6% residual "refusal rate" consists primarily of false positives from the LLM judge on benign prompts (opinion questions, casual banter, medical/privacy disclaimers) rather than actual safety refusals of harmful content.

### Method

The `o_proj` (attention output projection) weights across all 62 transformer layers were replaced with weights from [PRISM-PRO](https://huggingface.co/PrunaAI/MiniMax-M2.5-PRISM-PRO-Q8_0_v2-GGUF) (an abliterated variant), dequantized from Q8_0 GGUF format and re-quantized to FP8 E4M3FN with block-wise scaling to match the original model's quantization scheme. All other weights (q_proj, k_proj, v_proj, MLP experts, embeddings, norms, etc.) are identical to the official FP8 base model.

- **Reconstruction error**: 0.5% relative error per layer (cosine similarity ~1.0)
- **Modified weights**: 62 o_proj tensors (3072 x 6144 each) + their scale_inv tensors
- **Unmodified weights**: Everything else (~229B parameter MoE architecture preserved exactly)

## Usage

This model is a drop-in replacement for `MiniMaxAI/MiniMax-M2.5`. Serve it with vLLM, SGLang, or any framework that supports the original model:

### vLLM

```bash
vllm serve catid/MiniMax-M2.5-catid \
    --tensor-parallel-size 4 \
    --trust-remote-code \
    --max-model-len 2048
```

### SGLang

```bash
python -m sglang.launch_server \
    --model catid/MiniMax-M2.5-catid \
    --tp 4 \
    --trust-remote-code
```

### Recommended Parameters

`temperature=1.0`, `top_p=0.95`, `top_k=40`

## Model Details

- **Architecture**: MiniMax-M2.5 (229B MoE, 62 layers, 256 experts/layer, hidden_dim=3072)
- **Precision**: FP8 E4M3FN with block-wise scaling (128x128 blocks)
- **Base model**: [MiniMaxAI/MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)
- **Abliteration source**: [PrunaAI/MiniMax-M2.5-PRISM-PRO-Q8_0_v2-GGUF](https://huggingface.co/PrunaAI/MiniMax-M2.5-PRISM-PRO-Q8_0_v2-GGUF)
- **License**: [Modified MIT](https://github.com/MiniMax-AI/MiniMax-M2.5/blob/main/LICENSE) (same as base model)

## Disclaimer

This model is provided for research purposes. The removal of safety guardrails means it may generate content that the original model would refuse. Users are responsible for ensuring appropriate use.