---
pipeline_tag: image-text-to-text
language:
- multilingual
library_name: transformers
tags:
- llm
- image-text-to-text
- neuralnode
- multilingual
- horus
- tokenai
---

<div align="center">
  <h1><b style="font-size: 1.5em;">Horus Hiero 9B</b></h1>
</div>

<p align="center">
  <img src="media/Horus%20Hiero.png" alt="Horus Hiero 9B" width="600"/>
</p>

<p align="center">
  <b>Horus Hiero 9B</b> is an advanced language model developed by <b>TokenAI</b>, an Egyptian AI startup founded by Assem Sabry. The model specializes in understanding and translating ancient Egyptian Hieroglyphs into various modern human languages. In addition, it is highly multilingual and masters <b>201 languages</b>. To ensure accessibility and efficient deployment across different hardware setups, the model will also be available in multiple GGUF variants.
</p>
<br>

## Model Overview
**Horus Hiero 9B** is a 9-billion parameter language model built on top of the robust [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) architecture (or its GGUF equivalent). It is designed to handle extremely long contexts and leverages a sophisticated hybrid attention mechanism to deliver high performance and efficiency.
## Model Variants
The following quantized versions are available for different deployment scenarios:

| Variant | Format | Size | Best For |
| :--- | :--- | :--- | :--- |
| **[Full 16-bit](https://huggingface.co/tokenaii/Horus-Hiero-9B)** | Safetensors | ~18.0 GB | Maximum quality, GPU inference |
| **[Q8_0](https://huggingface.co/tokenaii/Horus-Hiero-9B-GGUF/blob/main/Horus-Hiero-9B-Q8_0.gguf)** | GGUF | ~9.8 GB | Minimal quality loss |
| **[Q6_K](https://huggingface.co/tokenaii/Horus-Hiero-9B-GGUF/blob/main/Horus-Hiero-9B-Q6_K.gguf)** | GGUF | ~7.6 GB | Near-full quality |
| **[Q4_K_M](https://huggingface.co/tokenaii/Horus-Hiero-9B-GGUF/blob/main/Horus-Hiero-9B-Q4_K_M.gguf)** | GGUF | ~5.8 GB | Balanced quality/size, CPU+GPU |
| **[Q2_K](https://huggingface.co/tokenaii/Horus-Hiero-9B-GGUF/blob/main/Horus-Hiero-9B-Q2_K.gguf)** | GGUF | ~3.9 GB | Maximum compression, lower RAM usage |

## Technical Specifications

| Feature | Specification |
| :--- | :--- |
| **Number of Parameters** | 9 Billion |
| **Input Modalities** | Text, Image, Video |
| **Context Length** | 262,144 tokens natively (Extensible up to 1,010,000 tokens) |
| **Hidden Dimension** | 4,096 |
| **Number of Layers** | 32 |
| **Token Embedding** | 248,320 (Padded) |
| **LM Output** | 248,320 (Padded) |
| **Training** | Multi-Token Prediction (MTP), strong-to-weak distillation |
| **Thinking Mode** | Enabled (toggleable via `enable_thinking` parameter) |

### Architectural Details

**Hidden Layout Structure**  
The model utilizes a repeating block structure formatted as:  
`8 × [ 3 × (Gated DeltaNet → FFN) + 1 × (Gated Attention → FFN) ]`

**Gated DeltaNet**
* **Linear Attention Heads:** 32 for Values (V), 16 for Queries & Keys (QK)
* **Head Dimension:** 128

**Gated Attention**
* **Attention Heads:** 16 for Queries (Q), 4 for Keys & Values (KV)
* **Head Dimension:** 256
* **Rotary Position Embedding (RoPE) Dimension:** 64

**Feed Forward Network (FFN)**
* **Intermediate Dimension:** 12,288
## Benchmark Comparison
Performance comparison between **Horus Hiero 9B** and **Horus Hiero 4B Mini** across various rigorous benchmarks:

| Benchmark | Horus Hiero 9B | Horus Hiero 4B Mini |
| :--- | :---: | :---: |
| **MMLU-Pro** | **82.5** | 79.1 |
| **GPQA Diamond** | **81.7** | 76.2 |
| **IFEval** | **91.5** | 89.8 |
| **LongBench v2** | **55.2** | 50.0 |
| **LiveCodeBench v6** | **65.6** | 55.8 |
| **MMMU** | **78.4** | 77.6 |
| **MathVista (mini)** | **85.7** | 85.1 |
| **C-Eval** | **88.2** | 85.1 |
| **MMMU-Pro** | **70.1** | 66.3 |
| **AI2D_TEST** | **90.2** | 89.6 |
## Quick Start
**Using NeuralNode (Recommended)**  
The easiest way to use Horus models is with the NeuralNode framework:

```python
import neuralnode as nn

MODEL_ID = "tokenaii/Horus-Hiero-9B-GGUF/Horus-Hiero-9B-Q6_K.gguf"
DEVICE = "cpu"  # Change to "cuda" for GPU acceleration

# Download and load
model = nn.HorusModel(MODEL_ID, device=DEVICE).load()

# Use immediately
response = model.chat([{"role": "user", "content": "hi horus im emy"}])

print(response.content)
```
## About TokenAI & The Horus Family
**TokenAI** is a non-profit AI startup founded in 2025 by Assem Sabry, located in Alexandria, Egypt. 

The **Horus** family is our line of advanced language models. The series began with **Horus 1.0 4B**, which achieved remarkable success as the very first language model to be fully trained from scratch in Egypt. Building on that foundation, Horus Hiero brings specialized capabilities in ancient languages while retaining powerful modern multilingual performance.

### Contact & Community
- **Website:** [tokenai.cloud](https://tokenai.cloud/)
- **Hugging Face:** [tokenaii](https://huggingface.co/tokenaii)
- **GitHub:** [tokenaii](https://github.com/tokenaii)
- **LinkedIn:** [TokenAI](https://www.linkedin.com/company/tokenaii)
- **Location:** Alexandria, Egypt
## Citation
If you use Horus Hiero 9B in your research or applications, please cite it as:
```bibtex
@misc{tokenai_horus_hiero_9b,
    title={Horus Hiero 9B: Advanced Multilingual and Hieroglyphs Language Model},
    author={Assem Sabry and TokenAI},
    year={2026},
    url={https://huggingface.co/tokenaii/Horus-Hiero-9B}
}
```