Korean startup Moreh tops global large language model test

South Korean artificial intelligence tech startup Moreh Inc. said on Thursday that its large language model (LLM) topped a performance assessment by global leading AI platform operator Hugging Face Inc.

The competition, Open LLM Leaderboard, measures the effectiveness of generative language models based on six benchmarks such as science questions, commonsense inference and mathematical reasoning. Moreh said it has ranked top with a score of 77.29.

The evaluation tracks an AI model’s capabilities in terms of knowledge, reasoning and math in various scenarios. There are more than 3,400 models from global AI companies on the platform, Moreh said.  

Moreh has trained the LLM with 70 billion parameters, dubbed MoMo-70B, by using the startup’s in-house platform MoAI and US chipmaker Advanced Micro Devices Inc.’s Instinct MI250 accelerator.

MoAI can maximize the efficiency of machine learning by encapsulating a large graphics processing unit (GPU) cluster system as a single device, Moreh said.

Moreh’s victory comes after its crosstown rival Upstage took the top spot.

Solar, Upstage’s LLM with 10.7 billion parameters, took a top position at Open LLM Leaderboard last month with a 74.2 score. Last August, Upstage’s AI model trained with Meta’s LLaMA 2 70B model rose to an average score of 72.3 on the platform. 

Write to Kang-Ho Jang at callme@hankyung.com

Jihyun Kim edited this article.