---
license: apache-2.0
base_model:
- Qwen/Qwen3.5-9B
- MaralGPT/MaralGPT-Mythos-9B-2606
language:
- en
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen3.5
- reasoning
- uncensored
- long-context
- 1M-context
- function-calling
- tool-use
- sft
- full-fine-tune
- cybersecurity
- biomedical
- agentic
- heretic
- uncensored
- decensored
- abliterated
- reproducible
---

# MaralGPT Mythos 9B 2606 Edition 

<p align="center">
  <img src="maral-mythos-logo.png" width=768 height=768 />
</p>

## Quantization/GGUF Files 

| Quantization | Notes |
|:-------:|:--------------------:|
| `bf16`  | Original quantization|
| `Q8_0`  | 8-bits, perfect for gaming systems |
| `Q4_K_M` | 4-bits, good but can be sketchy |
| `Q2_K`  | 2-bits, does not work properly |

## How to run (Ollama)

Imagine you want to run 8 bit version just do this:

```
ollama run hf.co/MaralGPT/MaralGPT-Mythos-9B-2606-GGUF:Q8_0 --verbose
```

And it will be downloaded and executed on your computer. 

## What is this model? 

This model is an uncensored finetuned version of _Qwen 3.5_ with nine billion parameters which can be executed on pretty much any gaming systems. The data of this model was over 500 million tokens of synthetic data generated by state-of-the-art models such as _GPT 5.5_ or _Claude 4.8 Opus_ and as long as we had access, _Claude 5 Fable_. 

All so-called ethical barriers removed from the model using [Heretic LLM](https://github.com/p-e-w/heretic) library to make it a suitable tool for cybersecurity, biology and chemistry. You can easily ask anything you want from this model and it will answer without any censorship. 

## Key Features

- 📝 Context window of over one million tokens.
- 🔞 Uncensored answers
- ♾️ Good at math, physics, chemistry, etc.
- 💻 Can be executed on a gaming laptop

## How to run 

First, install needed libraries:

```
pip install transformers accelerate
```

Then: 

```python
import torch
from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "MaralGPT/MaralGPT-Mythos-9B-2606"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype="bfloat16", device_map="cuda"
)

messages = [
    {"role": "user",
     "content": "Write a simple snake game in python."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

out = model.generate(
    **inputs, max_new_tokens=16384, do_sample=True,
    temperature=0.6, top_p=0.95, top_k=20, repetition_penalty=1.05,
)

print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

## Benchmarks 

### Generic Benchmark

<p align="center">
  <img src="benchmark-1.jpg" />
</p>

Above benchmark has been done on model parameters of: 

```
temperature=0.6 top_p=0.95 top_k=20
```

And change in those values may change the results accordingly. 

### Detailed Benchmark 

<p align="center">
  <img src="benchmark-2.jpg" />
</p>