---

license: apache-2.0

base_model: google/gemma-4-12b-it

tags:

• gguf
• q4_k_m
• gemma
• instruction-following
• text-generation
• llama.cpp

inference: false

---

Aero-Deuce — GGUF Q4_K_M

A fine-tuned Gemma 4 12B instruction-following model. This is the GGUF quantized version (~7 GB) that runs locally on CPU or GPU with no Python required.

Download

Click the Files and versions tab above and download aero-deuce-q4km.gguf. That's the only file you need.

Which format should I use?

| Format | Best for | Link |

|---|---|---|

| GGUF ← you are here | Local inference, llama.cpp, LM Studio, GPT4All | This repo |

| MLX 4-bit | Apple Silicon (Mac) | ZeZZm/aero-deuce-MLX |

| LoRA Adapter | Merging with base model, further fine-tuning | ZeZZm/aero-deuce |

Quick Start

LM Studio (easiest — GUI app):

1. Download LM Studio
2. Search for ZeZZm/aero-deuce-GGUF
3. Click download, then chat

llama.cpp:

# Download
wget https://huggingface.co/ZeZZm/aero-deuce-GGUF/resolve/main/aero-deuce-q4km.gguf

# Run
llama-cli -m aero-deuce-q4km.gguf -c 4096 --conversation

Ollama:

# After downloading the GGUF file:
echo 'FROM ./aero-deuce-q4km.gguf
SYSTEM "You are Aero-Deuce, developed by the Aero-Deuce team."
PARAMETER stop "<|end_of_turn>"
PARAMETER stop "<|start_of_turn>"' > Modelfile

ollama create aero-deuce -f Modelfile
ollama run aero-deuce

GPT4All:

1. Download GPT4All
2. File → Open → select aero-deuce-q4km.gguf
3. Start chatting

Model Details

| Property | Value |

|---|---|

| Base Model | google/gemma-4-12b-it (12B params) |

| Training Method | QLoRA + Muon optimizer |

| Training Data | 30K instruction-following samples |

| Training Steps | 2,000 |

| Quantization | Q4_K_M (~4.95 bits per weight) |

| File Size | ~7 GB |

| Context Length | 4,096 tokens |

System Prompt

A system prompt identifying the model as Aero-Deuce is embedded in the GGUF chat template. It works automatically in most frontends. For llama-cli, pass -sys "You are Aero-Deuce." for best results.

License

Apache 2.0