Refract AI Labs

Refract AI Labs is an experimental AI research group focused on efficient language-model architectures, compact reasoning systems, and unconventional alternatives to standard dense Transformers.

We build small models, weird architectures, and research prototypes with one main goal:

Maximum capability, minimum parameters.

News

28 September 2026 Refract AI is now a seperate group from smilyai!

Current projects

Orion / Project Prism

Orion is our experimental language-model family exploring:

• explicit temporary working state
• recurrent computation
• parameter reuse
• sparse conditional computation
• compact knowledge and procedure mechanisms
• hybrid base + instruction training

The current public model is:

Orion Flagship Mini
A ~219M parameter custom-architecture causal language model and instruction-tuned assistant.

Model: https://huggingface.co/Refract-Labs/Orion-Flagship-Mini

Base model: https://huggingface.co/Refract-Labs/Orion-Flagship-Mini-Base

Research direction

We are especially interested in:

• efficient LLM architectures
• recurrent and iterative reasoning
• alternatives to scaling parameter count
• parameter-efficient computation
• long-context behavior
• model memory and working state
• post-training and hybrid training methods
• small-model capability research

Philosophy

We like experiments that are:

• measurable
• reproducible
• honest about limitations
• interesting even when they fail

Architecture claims should be tested with controlled comparisons rather than assumed from design alone.

Open research

Some Refract AI Labs models and code are released publicly for research, experimentation, and community feedback.

Expect prototypes, unfinished ideas, occasional cursed training runs, and hopefully a few genuinely useful discoveries. 🧪

Part of the wider SmilyAI research ecosystem. (still kinda)