Afaf/atlas-mini

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:May 14, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

Afaf/atlas-mini is a 3.1 billion parameter Qwen2-based causal language model developed by Afaf, fine-tuned from unsloth/Qwen2.5-3B-Instruct-bnb-4bit. This model was trained using Unsloth and Huggingface's TRL library, enabling 2x faster training. It is designed for general instruction-following tasks, leveraging its efficient training methodology.

Loading preview...

Model Overview

Afaf/atlas-mini is a 3.1 billion parameter instruction-tuned language model developed by Afaf. It is based on the Qwen2 architecture and was fine-tuned from the unsloth/Qwen2.5-3B-Instruct-bnb-4bit model. A key differentiator of this model is its training methodology, which utilized Unsloth and Huggingface's TRL library, resulting in a 2x speedup during the training process.

Key Characteristics

  • Architecture: Qwen2-based, a robust causal language model.
  • Parameter Count: 3.1 billion parameters, offering a balance between performance and computational efficiency.
  • Training Efficiency: Leverages Unsloth for significantly faster fine-tuning.
  • License: Released under the Apache-2.0 license, allowing for broad use and distribution.

Use Cases

This model is suitable for a variety of general instruction-following tasks, benefiting from its efficient fine-tuning. Its relatively compact size makes it a good candidate for applications where computational resources are a consideration, while still providing solid performance for common NLP tasks.