~/LLM/motif-technologies-releases-motif-3-a-314b-mixture-of-experts-model

Motif-Technologies Releases Motif-3, a 314B Mixture-of-Experts Model

Motif-Technologies has officially announced Motif-3, a 314-billion parameter Mixture-of-Experts (MoE) model with 13 billion active parameters. Developed under South Korea's "Dokpamo" project, the model reportedly achieves competitive scores against larger models like Qwen-3.7-Max and DeepSeek-v4-Pro on agentic and coding benchmarks. This release highlights South Korea's growing competitiveness in the global AI landscape through its state-backed sovereign AI initiatives. It demonstrates that smaller, optimized MoE models can rival trillion-parameter models on complex reasoning and agentic tasks. Motif-3 scored 38.7 on GDPVal v2 and 76.2% on SWE-Bench Verified, placing it ahead of or on par with much larger models like GLM-5.1 and Kimi-K2.6. However, the announcement currently lacks detailed technical documentation and independent community validation.

## BACKGROUND

"Dokpamo" is South Korea's national initiative to build independent, sovereign AI foundation models to reduce reliance on foreign tech giants. Mixture-of-Experts (MoE) is an AI architecture that activates only a subset of its parameters (active parameters) for each input, improving efficiency compared to dense models.

## REFERENCES

## KEYWORDS

#LLM#Mixture of Experts#AI Benchmarks#Machine Learning

$ subscribe --daily

Motif-Technologies Releases Motif-3, a 314B Mixture-of-Experts Model | Daily News