✦
Compact but capable
Designed for efficient deployment without sacrificing creative depth or output quality.
Proof-of-concept • MacBook Air M4 showcase
Unrealistic-v1 is a compact AI proof-of-concept built to showcase M4 Base Air performance.
A 190M-parameter decoder-only transformer trained from scratch on a MacBook Air M4 with MLX, demonstrating what a compact, locally-run model can do when paired with Apple Silicon efficiency and modern training pipelines.
- 190MParameters
- M4Base Air
- Apache 2.0License
Model status
U
Unrealistic-v1
190M
Creative AI
Research
Generation
Efficiency
Prototype
Why it stands out
Built for performance, flexibility, and ambitious builders.
▣
Fast experimentation
Ideal for rapid iteration, local workflows, and prototyping ideas at a practical scale.
◎
Creative control
Suitable for generative and exploratory tasks where style, nuance, and output quality matter.
⬢
Production-minded
A model that balances portability and power for builders who want real-world usability.
Design philosophy
High quality output from a lean architecture.
Unrealistic-v1 was created with a focus on streamlined learning, expressive generation, and practical deployment. It brings together compact parameter efficiency with ambitious creative potential.
The model is strongest in general knowledge, grounded factual recall, and compact local execution—while still being honest about its limits in large-number word problems, deeper reasoning, biology, and complex multi-step tasks.
Model class
Compact foundation
Target
Generative workflows
Core value
Performance + honesty
Strengths and limits
Credible capability, without the hype.
Power
- Strong general knowledge and factual recall
- Efficient local inference on Apple Silicon
- Compact deployment profile for experimentation
- Good fit for compact research and creative tooling
Weakness
- Large-number word problems remain difficult
- Deep multi-step reasoning is limited
- Biology and complex programming tasks are weaker
- Not positioned as a frontier-class reasoning model
Potential applications
Built for modern AI workflows.
Creative tooling
Assist with ideation, concept generation, and expressive content work.
Research prototyping
Explore model behavior in a lightweight environment suitable for experimentation.
Product workflows
Power intelligent features without the heavy resource cost of large models.
Launch it