---
base_model: unsloth/Ministral-3-8B-Instruct-2512
tags:
- unsloth
- mistral
- gguf
- coding
- software-engineering
- rs.ai
license: apache-2.0
language:
- en
pipeline_tag: text-generation
---

# Scar-8B 💻

**Scar** is an elite, 8-billion parameter software engineering AI developed by **rs.ai**. 

Specifically engineered as a universal coding agent, Scar possesses deep capabilities across all major programming languages, system architectures, multi-agent frameworks, and backend workflows. It outputs highly optimized, production-ready code with clean markdown formatting.

## Model Details
- **Developer:** rs.ai 
- **Base Model:** Ministral-3-8B-Instruct
- **Parameters:** 8 Billion
- **Quantization:** GGUF (`Q4_K_M`) and 16-bit LoRA adapters
- **Training Framework:** Unsloth (PEFT/LoRA)
- **Primary Use Case:** Agentic AI pipelines, backend engineering, code generation, and technical reasoning.

## How to Use

### Python Inference (Transformers / Unsloth)
You can run the model natively in Python using `transformers` or `unsloth`.

```python
from unsloth import FastLanguageModel
from transformers import TextStreamer

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "ItzRustam/Scar-8B",
    max_seq_length = 8192,
    dtype = None,
    load_in_4bit = True,
)

FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are Scar, an elite software engineering AI developed by rs.ai. You possess absolute mastery over all programming languages, system architectures, and backend frameworks."},
    {"role": "user", "content": "Write a high-performance HTTP server in Go."}
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
text_streamer = TextStreamer(tokenizer, skip_prompt=True)

_ = model.generate(input_ids=inputs, streamer=text_streamer, max_new_tokens=512, temperature=0.5)
```

## Training Details
Scar-8B was fine-tuned using a specialized, ShareGPT-formatted dataset featuring **Compound Intent Resolution**. This ensures the model can accurately process complex, multi-step system prompts—such as simultaneously defining its internal identity and generating complex code blocks—without suffering from early sequence truncation (`[EOS]` drop-offs).

## License & Attribution
This model is published under the permissive **Apache 2.0 License**, allowing for commercial and private deployment. 

*Base Architecture Attribution:* Scar-8B is a derivative model post-trained on top of the `Ministral-3-8B-Instruct-2512` architecture originally developed by Mistral AI.