ItzRustam/Scar-8B-GGUF

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 also run the model natively in Python using transformers or unsloth.

from unsloth import FastLanguageModel
from transformers import TextStreamer

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "ItzRustam/Scar-8B-GGUF",
    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.