Docs
Everything needed to understand the model and its intent.
Unrealistic-v1 is presented as a research and proof-of-concept model: a compact, locally-trainable transformer that demonstrates the strengths of MacBook Air M4 hardware in a practical AI workflow.
Overview
Unrealistic-v1 is a 190M-parameter decoder-only transformer trained from scratch on a MacBook Air M4 using MLX. It is designed as a compact, efficient model for experimentation, local inference, and creative AI prototyping.
Hardware showcase
The model is explicitly used to demonstrate the M4 Base Air platform: 8-core GPU, 10-core CPU, and 16-core Neural Engine performance in a practical foundation-model workflow.
Licensing
License: Apache 2.0. The model and its surrounding materials are distributed under the Apache License 2.0, with code and generated content use subject to the terms of that open-source license.
Quick FAQ
What is Unrealistic-v1?
A proof-of-concept transformer model designed to showcase local MLX training and inference on a MacBook Air M4 system.
Why is it called a PoC?
Because it acts as a compact research artifact and demonstration of what a lightweight, locally-run model can achieve on Apple Silicon hardware.
What is it good at?
General knowledge recall, compact local inference, and practical experimentation with small-scale training.
What are its weaknesses?
It remains limited in deep reasoning, large-number word problems, biology, and complex multi-step tasks, which is why the project presents itself honestly as a proof of concept.
What license does it use?
The project uses MIT for code, Apache 2.0 for model weights, and CC BY 4.0 for documentation. Data sources are also documented in the repository.
Where can I access it?
Open the model directly on Hugging Face using the official model page, and review the included training and evaluation docs in the repository.
Licensing & data