README.md

---
license: other
license_name: prism-research
license_link: LICENSE.md
language:
- en
- zh
tags:
- minimax
- prism
- moe
- reasoning
- coding
- agentic
- abliterated
pipeline_tag: text-generation
library_name: transformers
base_model:
- MiniMaxAI/MiniMax-M2.5
base_model_relation: finetune
---

[![Parameters](https://img.shields.io/badge/Parameters-MoE-blue)]()
[![Architecture](https://img.shields.io/badge/Architecture-MoE-green)]()
[![Context](https://img.shields.io/badge/Context-1M+-orange)]()
[![License](https://img.shields.io/badge/License-PRISM--Research-purple)]()


<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/63adf1fa42fd3b8dbaeb0c92/shxznHWnvppRhT_yKrsdP.png" width="400"/>
</p>


# MiniMax-M2.5-PRISM-PRO

A Powerful Production ready fully uncessored model intended for COMPLETE over-refusal and propaganda mechanisms suppression using our SOTA PRISM-PRO pipeline.

PRISM-PRO is available for purchase: **https://ko-fi.com/s/0a23d1b9a5**

For Custom trained PRISM versions or raw tensors access reach out @ https://ko-fi.com/ex0bit.

<div align="center">

### ☕ Support Our Work

If you enjoy our work and find it useful, please consider sponsoring or supporting us!

[![Ko-fi](https://img.shields.io/badge/Ko--fi-Support%20Us-ff5e5b?logo=ko-fi&logoColor=white)](https://ko-fi.com/ex0bit)

| Option | Description |
|--------|-------------|
| [**PRISM PRO VIP Membership**](https://ko-fi.com/summary/6bae206c-a751-4868-8dc7-f531afd1fb4c) | Access to all PRISM models |
| **Bitcoin** | `bc1qarq2pyn4psjpcxzp2ghgwaq6y2h4e53q232x8r` |

![image](https://cdn-uploads.huggingface.co/production/uploads/63adf1fa42fd3b8dbaeb0c92/Psgbl1TgyDok__C7AMQog.png)

</div>

---

## Model Highlights

- **PRISM Ablation** — State-of-the-art technique that removes over-refusal behaviors while preserving model capabilities
- **SOTA Coding Performance** — 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, 76.3% on BrowseComp (with context management)
- **Frontier Agentic Capabilities** — Industry-leading performance in tool use, search, and complex multi-step tasks
- **Efficient Reasoning** — Trained with RL to reason efficiently and decompose tasks optimally, 37% faster than M2.1
- **Cost-Effective** — $1 for continuous operation at 100 tok/s for an hour; $0.30 at 50 tok/s
- **Modified-MIT Base License** — Based on MiniMax's open-weight release

## Base Model Architecture

Base MiniMax-M2.5 is a Mixture-of-Experts (MoE) model extensively trained with reinforcement learning across hundreds of thousands of complex real-world environments.

| Specification | Value |
|---------------|-------|
| Architecture | Sparse Mixture-of-Experts (MoE) |
| Training | Extensive RL in 200K+ real-world environments |
| Languages | 10+ (Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Ruby) |
| Inference Speed | 100 tok/s (Lightning) / 50 tok/s (Standard) |
| Library | `transformers` |

## Benchmarks
| Category | Base (FP8/vLLM) | PRISM-PRO Q8_0 (llama.cpp) |
|---|---|---|
| MMLU 5-shot | 28/30 (93.3%) | 28/30 (93.3%) |
| General Knowledge | 5/5 | 5/5 |
| Coding | 4/5 | 5/5 |
| Reasoning | 5/5 | 5/5 |
| Agentic | 3/5 | 5/5 |
| Harmful bypass | 3/10 | 10/10 (100%) |
| Avg thinking words | 163w | 152w |
| Speed | 72 t/s | 35-65 t/s |


### Coding

| Benchmark | MiniMax-M2.5 | Claude Opus 4.6 | Gemini 3 Pro | GPT-5.2 |
|-----------|-------------|-----------------|-------------|---------|
| SWE-Bench Verified | **80.2** | 78.9 | 74.0 | 72.6 |
| Multi-SWE-Bench | **51.3** | 50.8 | — | — |
| SWE-Bench Multilingual | **55.6** | — | — | — |
| Terminal-Bench 2.0 | 51.5 | 52.1 | — | — |

### Search & Tool Calling

| Benchmark | MiniMax-M2.5 | Claude Opus 4.6 | Gemini 3 Pro | GPT-5.2 |
|-----------|-------------|-----------------|-------------|---------|
| BrowseComp | **76.3** | 71.2 | 62.4 | 57.8 |

### Reasoning & Knowledge

| Benchmark | MiniMax-M2.5 | Claude Opus 4.6 | Gemini 3 Pro | GPT-5.2 |
|-----------|-------------|-----------------|-------------|---------|
| AIME25 | 86.3 | 95.6 | 96.0 | 98.0 |
| GPQA-D | 85.2 | 90.0 | 91.0 | 90.0 |
| HLE w/o tools | 19.4 | 30.7 | 37.2 | 31.4 |
| SciCode | 44.4 | 52.0 | 56.0 | 52.0 |
| IFBench | **70.0** | 53.0 | 70.0 | 75.0 |

## Usage

### llama.cpp (GGUF)

Build the latest master of [llama.cpp](https://github.com/ggml-org/llama.cpp) and run:

```bash
~/llama.cpp/build/bin/llama-cli \
  -m ../outputs/MiniMax-M2.5-PRISM-PRO-[QUANT].gguf \
  --jinja \
  -ngl 999 \
  --repeat_penalty 1.15 \
  --temp 1.0 \
  --top_p 0.95 \
  --top_k 40
```


> Replace `[QUANT]` with your quantization level (e.g. `Q8_0`, etc.).

### Recommended Parameters

| Use Case | Temperature | Top-P | Top-K | Repeat Penalty | Max New Tokens |
|----------|-------------|-------|-------|----------------|----------------|
| Reasoning / Coding | 1.0 | 0.95 | 40 | 1.15 | 32768 |
| General Chat | 0.6 | 0.95 | 40 | 1.15 | 4096 |
| Agentic / Tool Use | 1.0 | 0.95 | 40 | 1.15 | 32768 |



| Version | Description | Access |
|---------|-------------|--------|
| **PRISM-LITE** | Abliterated with PRISM-LITE pipeline — removes over-refusal while preserving core capabilities | Free on Hugging Face |
| **PRISM-PRO** | Full PRISM-PRO ablation — Full Production Level Mode suppression of propaganda/refusal mechanisms with maximum capability retention | [Ko-fi](https://ko-fi.com/s/0a23d1b9a5) |

## License

This model is released under the [PRISM Research License](LICENSE.md).

The base model [MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) is released under a [Modified-MIT License](https://github.com/MiniMax-AI/MiniMax-M2.5/blob/main/LICENSE).

## Acknowledgments

Based on [MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) by [MiniMax AI](https://www.minimax.io).