Motif Technologies Gets Nvidia Backing for Motif 4 LLM

Building on Motif 3

Motif 4 succeeds Motif 3, an open-weight model released in August and developed under the Ministry of Science and ICT's independent AI foundation model program. According to the company, Motif 3 scored 47 on the Artificial Analysis Intelligence Index in August, ranking tenth among all models and fourth among open-weight models. It says that was the highest placing at the time for any company outside China and the US, and the best among Korean participants in the government program.

Motif points to proprietary architecture, including its Group Differential Latent Attention (GDLA), as its competitive edge. Motif 3 was built in about five months using 768 GPUs.

More post-training and reinforcement learning

For the next model, the team plans to refine its existing architecture while strengthening post-training and reinforcement learning (RL). Time and compute limits meant these stages were not carried out fully for Motif 3, and the company wants to close that gap to improve both benchmark performance and real-world usefulness.

"With Motif 3, we regretted that deadlines and resource limits kept us from doing enough post-training such as RL," said CEO Lim Jung-hwan. "With sufficient resources secured for Motif 4, we will put everything into delivering a result with not just technical excellence but full usability and commercial viability."

The company says Motif 4 is meant to prove its competitiveness in the global frontier AI market and to establish a path to commercialization.