Will Korean Version of Llama 3 Be Released... Korean Startup Proposes to Zuckerberg

Upstage "Proposal for Developing Versions in Various Languages"
Enhancing Language Diversity and Accuracy Enables Model Differentiation

South Korean artificial intelligence (AI) startup Upstage has reportedly met with Mark Zuckerberg, CEO of Meta, to propose the development of localized versions of the next-generation large language model (LLM) ‘LLaMA 3’ in Korean and other languages. Zuckerberg responded positively, saying he would consider the proposal. Unlike LLMs released by major big tech companies such as OpenAI, which have trained Korean language models with limited data, the intention behind ‘LLaMA 3’ is to create separate versions for each country. If accepted, this approach is expected to enhance information accuracy and enable targeting each market with differentiated models.

Mark Zuckerberg, CEO of Meta, is arriving in Seoul through the Seoul Gimpo Business Aviation Center in Gangseo-gu on the 27th. Visiting Korea for the first time in 10 years, CEO Zuckerberg is expected to meet President Yoon Suk-yeol and discuss AI semiconductor cooperation plans with Samsung Electronics Chairman Lee Jae-yong and LG Electronics CEO Cho Joo-wan during his 1-night, 2-day stay. Photo by Hyunmin Kim kimhyun81@

According to the IT industry on the 29th, about ten representatives and officials from domestic startups including Upstage and Devs United Games held a private meeting with Zuckerberg the day before. Kim Seong-hoon, CEO of Upstage, who attended the meeting, presented the idea of creating localized versions of LLaMA 3 to Zuckerberg. Tests of previous models showed that Korean language performance was lacking. However, due to the reality of insufficient Korean language data and related resources, he proposed a collaborative approach by selecting representative companies or developers from each country. It is reported that Zuckerberg responded positively to this proposal and said he would consider it.

An Upstage official said, "We suggested opening the Korean and Japanese versions of LLaMA 3 together to create a global impact," adding, "We look forward to collaborating on LLaMA 3."

If LLaMA 3 is released in localized language versions, it will have a distinctive advantage in dominating the global AI ecosystem. Most companies, including OpenAI and Google, develop LLMs based on English. However, when models are used in non-mainstream languages or languages with insufficient data, accuracy often declines.

It is reported that LLaMA 3 has completed internal development. Strategic decisions regarding the model’s release method remain, but there is growing expectation that it will be unveiled within this year.