---
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
library_name: peft
pipeline_tag: text-generation
tags:
- base_model:adapter:Qwen/Qwen2.5-Coder-3B-Instruct
- lora
- sft
- transformers
- trl
- german
- english
- coding
- code-generation
- aether
- gguf
license: apache-2.0
language:
- de
- en
---

# Aether 2.5 Coder

Aether 2.5 Coder is a specialized coding and technical reasoning model built on top of the **Qwen2.5-Coder-3B-Instruct** base architecture. Fine-tuned using SFT (Supervised Fine-Tuning) with Hugging Face TRL and PEFT (LoRA) on custom datasets, Aether 2.5 Coder combines high-precision code generation, script optimization, and debugging with multilingual instruction following in German and English.

> 🚀 **Looking for GGUF versions?**
> If you want to run Aether 2.5 Coder locally via **LM Studio**, **Ollama**, or **llama.cpp**, check out the pre-quantized GGUF repository:
> 👉 **[Maxilicious20/Aether-2.5-Coder-GGUF](https://huggingface.co/Maxilicious20/Aether-2.5-Coder-GGUF)**

## Model Details

### Model Description

- **Developed by:** Maxilicious20
- **Model type:** Causal Language Model (LoRA Adapter)
- **Language(s) (NLP):** German, English, Programming Languages (Python, JavaScript, C++, Luau, etc.)
- **License:** Apache-2.0
- **Finetuned from model:** Qwen/Qwen2.5-Coder-3B-Instruct

## Uses

### Direct Use

Aether 2.5 Coder is tailored for automated code completion, script writing, structural refactoring, debugging, and software architecture planning. It delivers top-tier 3B coding performance while maintaining low VRAM consumption for efficient execution on consumer hardware.

### Quantized & GGUF Models

For standalone CPU/GPU local execution without Python/Transformers dependencies, use the quantized GGUF binaries:

* 📦 **GGUF Repository:** [Maxilicious20/Aether-2.5-Coder-3B-GGUF](https://huggingface.co/Maxilicious20/Aether-2.5-Coder-3B-GGUF)
* **Available Quantizations:**
  * `aether_coder_f16.gguf` (Uncompressed / Full Precision)
  * `aether_coder_q8_0.gguf` (High Quality / 8-bit)
  * `aether_coder_q4_k_m.gguf` (Recommended / Balanced Speed & VRAM)

### How to Get Started with the Model

#### Python (Transformers & PEFT)

Use the following Python code to load Aether 2.5 Coder with `transformers` and `peft`:

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen2.5-Coder-3B-Instruct"
adapter_id = "Maxilicious20/Aether-2.5-Coder-3B"

# Load Tokenizer and Base Model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load Aether 2.5 Coder LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_id)

# Example Prompt
messages = [
    {"role": "system", "content": "You are Aether 2.5 Coder, an expert AI programming assistant."},
    {"role": "user", "content": "Write a Python script to filter and parse a JSON dataset efficiently."}
]

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

outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))