Motif 3 (Beta)

Motif 3 (Beta) is a large-scale Mixture-of-Experts (MoE) causal language model developed by the South Korean startup Motif Technologies. Released as part of South Korea's sovereign AI foundation model project, the model was built from the ground up using a proprietary architecture rather than re-parameterizing existing open-source frameworks. It features approximately 314.8 billion total parameters, though it maintains high computational efficiency during inference by selectively activating roughly 13 billion parameters per token.

The model's architecture introduces several novel components, most notably Grouped Differential Latent Attention (GDLA) and Grouped PolyNorm activation applied per expert. Its sparse routing mechanism utilizes 384 experts, with eight experts routed per token in addition to one shared expert. This design allows the model to handle complex reasoning tasks across a native context window of 262,144 tokens (256K), while remaining optimized for both Korean and English linguistic nuances.

In performance benchmarks, Motif 3 (Beta) achieved a score of 44 on the Artificial Analysis Intelligence Index (AAII), placing it among the highest-ranked open-weight models globally and ranking it as a leading model developed outside the United States and China. This beta version was released under a custom non-commercial license for personal and research use, representing an interim milestone before a final production release.