Motif Technologies' in-house AI foundation model Motif 3 scored the highest among models from South Korean companies on the Aggregate Intelligence Index (AAII) by global AI model evaluator Artificial Analysis.

Motif Technologies said on Aug. 13 that the official Motif 3 model scored 47 on AAII and that it released the model, training code and other materials as open source on Hugging Face.

AAII is currently the most heavily weighted external benchmark in the second-stage evaluation of the government's Independent AI Foundation Model (Dokpamo) project. Of the 100 points in the second-stage evaluation, benchmark assessment accounts for 40 points, and 25 points of that are allocated to AAII.

The company said Motif 3's AAII score of 47 ranks ninth among all global LLMs and fourth among open-weight models.

Motif 3 is a mixture-of-experts (MoE) large language model (LLM) that activates 13.2 billion parameters out of a total 314 billion. Motif Technologies developed it from scratch over about five months, from pre-training through post-processing.

A technical report released by Motif Technologies said Motif 3 applies its proprietary attention structure, Grouped Differential Latent Attention (GDLA), and was pre-trained using about 1.25 trillion tokens.

The company said it also delivered performance competitive with major overseas open-weight models on practical benchmarks.

On τ³-Banking, which evaluates AI agents' task performance in a financial work environment, it scored 35.3. On Terminal-Bench 2.1, which measures coding and system-task capability in a terminal environment, it scored 74.9.

It scored 76.2 on SWE-bench Verified, a software engineering performance evaluation, and received 37.0 on Humanity's Last Exam, a high-difficulty knowledge evaluation.

Motif Technologies also released the Motif 3 model weights, along with training code and libraries, under the MIT license. Companies, research institutions and developers can use the model for commercial purposes as well.

In addition to the language model, it also developed modules of 7 billion parameters for vision, 0.7 billion for audio and 1.2 billion for a video VAE. It plans to combine these modules with the language model to expand into a multimodal model.

Motif Technologies CEO Junghwan Lim (임정환) said, "Motif 3 is a starting point that proved we can compete globally with our own architecture alone in a short period and with limited resources." He added, "Now is the time to prove how far we can go when given the same conditions."