01
Lean footprint
Designed to operate with practical memory and compute requirements.
Model profile
Proof-of-concept intelligence built for the M4 Base Air era.
Unrealistic-v1 is a compact transformer model designed as a research and demonstration artifact: 190M parameters, trained from scratch on a MacBook Air M4 with MLX, and tuned to showcase the power of Apple Silicon in a small, capable local deployment profile.
190M
Parameters
8-core
GPU
10-core
CPU
16-core
Neural Engine
Core architecture
Optimized for quality without excess.
The model is deliberately designed around efficient scale, making it approachable for creators and researchers who want a powerful, portable model without demanding massive infrastructure.
It follows a LLaMA-family decoder-only transformer structure with 14 layers, hidden size 768, FFN size 4096, RoPE θ=10k, RMSNorm, SiLU/SwiGLU, and a 1024-token context window, trained from scratch on a MacBook Air M4 with MLX.
- Strong general knowledge and local inference efficiency
- Compact footprint suited to Apple Silicon workflows
- Known weaknesses in deep reasoning, biology, and multi-step tasks
- Designed as a serious proof-of-concept, not an overhyped replacement for large frontier models
Architecture
Compact transformer stack with a clear training path.
Token input
→
→
14 decoder blocks
RMSNorm · RoPE · Self-Attn · SwiGLU MLP
→
Final norm
→
Language head
Evaluation
Benchmark snapshot from the final checkpoint.
Model strengths
What gives Unrealistic-v1 its edge.
02
Expressive output
Balances compact scale with strong creative and generative behavior.
03
Easy iteration
Useful for real-world experimentation and rapid product discovery.