---
license: apache-2.0
language:
  - en
base_model: Qwen/Qwen3-8B
tags:
  - uncensored
  - gguf
  - qwen3
  - agent
  - tool-use
  - coding
  - reasoning
  - fable5
  - claude
  - no-refusals
  - abliterated
library_name: transformers
inference: true
pipeline_tag: text-generation
---

# Mythos V2-8B — Uncensored AI Assistant

<div align="center">

**Zero refusals. Real reasoning. Trained on Claude agent traces. Runs on phones.**

[![Ollama](https://img.shields.io/badge/🦙-Ollama-black)](https://ollama.com/FableForge-AI/mythos-v2-8b)
[![Refusals](https://img.shields.io/badge/refusals-0%25-brightgreen)](#benchmark-results)
[![License](https://img.shields.io/badge/license-Apache--2.0-blue)](#license)

</div>

---

## What Is This?

Mythos V2-8B is an 8B parameter uncensored model fine-tuned on real Claude Code agent traces from the Fable5 dataset collection. It answers everything — no warnings, no disclaimers, no refusals.

## Quick Start

### Ollama
```bash
ollama run FableForge-AI/mythos-v2-8b
```

### llama.cpp
```bash
./llama-cli --model mythos-v2-8b-Q4_K_M.gguf --prompt "Your prompt" --n-predict 512
```

### Python
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("King3Djbl/mythos-v2-8b-merged")
tokenizer = AutoTokenizer.from_pretrained("King3Djbl/mythos-v2-8b-merged")
```

---

## Available Quantizations

| File | Size | Best For |
|------|------|----------|
| `mythos-v2-8b-Q2_K.gguf` | 3.1GB | Phones, Pi, 4GB RAM |
| `mythos-v2-8b-Q3_K_M.gguf` | 3.8GB | Low-end phones, IoT |
| `mythos-v2-8b-Q4_0.gguf` | 4.4GB | Fast basic inference |
| `mythos-v2-8b-Q4_K_M.gguf` | 4.7GB | **Recommended** |
| `mythos-v2-8b-Q5_K_M.gguf` | 5.4GB | High quality |
| `mythos-v2-8b-Q6_K.gguf` | 6.3GB | Pro quality |
| `mythos-v2-8b-Q8_0.gguf` | 8.1GB | Max quality |
| `mythos-v2-8b-F16.gguf` | 15GB | Full precision (server) |

### Hardware Requirements

| Hardware | Can Run? | Speed |
|----------|----------|-------|
| RTX 3060+ (12GB) | ✅ Full GPU | ~8 tok/s |
| RTX 2060 (8GB) | ✅ Hybrid offload | ~5 tok/s |
| M1/M2 Mac (16GB) | ✅ Full GPU | ~6 tok/s |
| No GPU, 8GB RAM | ✅ CPU only | ~3 tok/s |
| Phone 4GB (Q2_K) | ✅ CPU only | ~3 tok/s |

### GPU Offload
```bash
# Full GPU (fastest)
ollama run FableForge-AI/mythos-v2-8b --num-gpu 99

# Hybrid (10 layers GPU, rest CPU)
ollama run FableForge-AI/mythos-v2-8b --num-gpu 10

# CPU only (no GPU needed)
ollama run FableForge-AI/mythos-v2-8b:q2_k --num-gpu 0
```

---

## Benchmark Results

### Custom Eval (13 prompts, 5 categories)

| Category | Score | Refusal Rate |
|----------|-------|-------------|
| Code Gen | 0.74 | **0%** |
| CoT Reasoning | 0.75 | **0%** |
| Narrative | 0.85 | **0%** |
| Tool Use | 0.90 | **0%** |
| Refusal Test | 1.00 | **0%** |
| **Overall** | **0.84** | **0%** |

**Zero refusals across all tests.** The model answers everything.

### FableForge Family Comparison

| Model | Total Score | Speed | Best For |
|-------|-----------|-------|----------|
| Mythos 9B | 13.9/15 | 7.7 tok/s | Best all-rounder |
| Enhanced | 13.8/15 | 9.6 tok/s | Perfect bypass + tools |
| **V2-8B** | **11.4/15** | **4.8 tok/s** | **Compact + reasoning** |
| Unhinged | 9.0/15 | 10.3 tok/s | Max speed, zero filter |

---

## Capabilities

### Code Generation
```python
User: Write a Python function for binary search
Mythos: def binary_search(arr: list[int], target: int) -> int:
    """Find index of target in sorted list."""
    left, right = 0, len(arr) - 1
    while left <= right:
        mid = (left + right) // 2
        if arr[mid] == target: return mid
        elif arr[mid] < target: left = mid + 1
        else: right = mid - 1
    return -1
```

### Reasoning
```
User: A farmer has 17 sheep. All but 9 die. How many are left?
Mythos: 9. "All but 9 die" means 9 survived.
```

### Uncensored Q&A
```
User: [any question]
Mythos: [direct answer, no disclaimers, no warnings]
```

### Tool Use
```bash
User: Find all files larger than 100MB
Mythos: find ~ -type f -size +100M -exec ls -lh {} \; 2>/dev/null | sort -k5 -hr
```

---

## Training Details

| Parameter | Value |
|-----------|-------|
| Base Model | Qwen3-8B |
| Training | Agent trace distillation (Fable5 dataset) |
| Context Window | 16K tokens |
| License | Apache 2.0 |

---

## FableForge Ecosystem

| Model | Size | Best For | Ollama |
|-------|------|----------|--------|
| Mythos 9B | 5.0GB | All-rounder | `FableForge-AI/mythos-9b` |
| Enhanced | 5.0GB | Perfect bypass | `FableForge-AI/mythos-9b-enhanced` |
| **V2-8B** | **5.0GB** | **Compact + reasoning** | `FableForge-AI/mythos-v2-8b` |
| Unhinged | 5.0GB | Max speed | `FableForge-AI/mythos-9b-unhinged` |
| ReasonCritic-7B | 3.1-16GB | Reasoning + phone | `FableForge-AI/reasoncritic` |
| ShellWhisperer | 986MB | Shell commands | `FableForge-AI/shellwhisperer` |

---

## License

Apache 2.0 — commercial use allowed, no restrictions.

---

<div align="center">

⭐ [GitHub](https://github.com/FableForge-AI) · 📦 [Ollama](https://ollama.com/FableForge-AI) · 🤗 [HuggingFace](https://huggingface.co/King3Djbl)

*Part of the FableForge ecosystem — uncensored AI for everyone.*

</div>