A Korean startup has entered the top ranks of the global large language model (LLM) leaderboard for the first time in a while. Since the Hugging Face leaderboard was revamped into Season 2 last year, this is the first time a Korean company has made it into the top 10.
As of the 24th, Linkbricks’ subsidiary, Linkbricks HorizonAI, ranks 10th on the Hugging Face OpenLLM leaderboard with an average accuracy of 47.34%.
The company’s model, ‘Saxo/Linkbricks-Horizon-AI-Avengers-V1-32B’, is a fine-tuned version of ‘Qwen/Qwen2.5-32B-Instruct’. Using eight H100-80G GPUs, the model has undergone supervised fine-tuning (SFT) and direct preference optimization (DPO) on approximately 35% of its parameters to enhance its multilingual capabilities. The tokenizer remains the same as the base model without any word expansions.
Notably, the model utilizes 80 million news and Wikipedia corpus entries from various language groups. It has also been trained with cross-language learning data for Japanese, Korean, Chinese, and English, along with mathematical and logical reasoning datasets. This allows it to handle “cross-language augmentation processing” and solve “complex logical problems”.
Additionally, the model excels in high-dimensional analysis of unstructured data, such as customer reviews and social media posts. It has also been optimized for coding, writing, mathematics, and logical reasoning.
Linkbricks Horizon-AI’s Vision and Achievements Linkbricks Horizon-AI, the developer of this model, was founded in January with the goal of building a small language model (sLM) development platform. The company is chaired by Ji Yun Sung(Benedict Ji), a data scientist.
Since last year, Ji has been developing a Japanese small language model (sLM) through the HorizonAI platform, which serves as a language model development tool. His model has also ranked high in the Open JP-LLM leaderboard, which assesses the performance of Japanese LLMs.
Moreover, in November 2023, HorizonAI topped the domestic Open Ko-LLM leaderboard (Season 2), securing leading positions in South Korea, Japan, and the United States.
On the 24th, Ji announced that HorizonAI has expanded into an “agent development platform” and is now completed and set for release.
“Since the LLM market is already saturated, the industry trend is now shifting towards efficiently combining existing LLMs to create practical services,” he stated. “HorizonAI will serve as a tool for real-world applications by handling the inference aspect of LLMs.”
Simplifying AI Agent Development is particularly focused on making AI agent development easier. By leveraging multiple LLMs, the platform can facilitate the creation of tools like ‘Deep Research’.
Users can drag and drop LLM components like Lego blocks within the HorizonAI platform to create their own agent services.
“All the necessary functions are built into the platform, so users only need to combine the elements,” Ji explained.
This means applications can be developed without complex fine-tuning. Additionally, the company has filed three related patents.
The completed applications can be applied to kiosks, robots, and various other systems, further expanding their practical use.
AI Times open japanese llm leaderboard Open llm leaderboard