---
license: apache-2.0
base_model: Umranz/Hydra-Turbo
tags:
- gguf
- llama.cpp
- qwen35
- quantized
- text-generation
pipeline_tag: text-generation
---

# Hydra-Turbo-GGUF

This repository contains official GGUF quantized weights for [Umranz/Hydra-Turbo](https://huggingface.co/Umranz/Hydra-Turbo), a 9B parameter model based on the Qwen 3.5 architecture.

The GGUF files in this repository were converted using the latest build of `llama.cpp` with the `--no-nextn` conversion flag to ensure compatibility across local inference runtimes.

---

## Model Overview

- **Base Model**: Umranz/Hydra-Turbo
- **Architecture**: Qwen 3.5 (Hybrid Linear Attention + Full Attention)
- **Parameters**: 8.95B (9.0B label)
- **Context Length**: 262,144 tokens
- **Vocabulary Size**: 248,320
- **License**: Apache 2.0

---

## Available Files and Quantization Formats

| File Name | Quantization Type | File Size | Description |
| :--- | :--- | :--- | :--- |
| `hydra-turbo-q4_0.gguf` | Q4_0 | 4.95 GB | Legacy 4-bit quantization. Fast execution, low VRAM footprint. |
| `hydra-turbo-q4_k_m.gguf` | Q4_K_M | 5.24 GB | Recommended medium 4-bit quant. Balanced accuracy and memory usage. |
| `hydra-turbo-q5_k_m.gguf` | Q5_K_M | 6.02 GB | High-precision 5-bit quant. Reduced perplexity loss with minimal speed overhead. |
| `hydra-turbo-q8_0.gguf` | Q8_0 | 8.87 GB | Near-lossless 8-bit quantization. Maximum output quality. |

---

## Technical Specifications

- **Layers**: 32 Main Transformer Blocks (SSM / Gated Delta Net + Full Attention)
- **Embedding Length**: 4096
- **Feed Forward Dimension**: 12288
- **Attention Heads**: 16 (Key-Value Heads: 4)
- **Rotary Position Embedding (RoPE)**: Frequency Base 10,000,000
- **Layer Norm Epsilon**: 1e-6

---

## Prompt Template

Hydra-Turbo uses the standard Qwen Chat Template:

```text
<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Your query here<|im_end|>
<|im_start|>assistant