Upstage CEO: "Business Multimodal AI Launching in June... Superior to LLaMA and Gemini"

Upstage Holds Media Day
Previews Multimodal 'Solar DocVLM' Combining OCR and LLM Technologies
"Vision Language Model More Accurate Than Google and Meta"
On ChatGPT Ghibli Copyright Controversy: "Unfair Competition... Reasonable Compensation Needed"
Ahead of Presidential Election: "Urgent Need to Secure Data, Talent, and GPUs"

Domestic AI development startup Upstage will launch 'Solar DocVLM,' a multimodal AI for business that simultaneously processes various data types such as text, images, voice, and video, in June. According to Upstage, Solar DocVLM has higher accuracy than Meta's 'LLaMA 4 Scout' and Google's 'Gemini 2.5 Pro.'

On the morning of the 16th, at a media day event held at the Conrad Hotel in Yeouido, Yeongdeungpo-gu, Seoul, Upstage CEO Kim Seong-hoon stated, "We will accelerate AI transformation across industries in the global market through 'Document Parse,' which processes documents using our self-developed optical character recognition (OCR) technology, and 'Solar,' our large language model (LLM)."

He added, "Companies like Upstage that have independently developed a full-stack AI model from OCR to LLM are rare worldwide." AI full-stack refers to encompassing foundational technologies such as semiconductors and cloud computing, as well as AI application services provided to users.

Upstage also emphasized its unique technological competitiveness in the area of industry-specific specialized small language models (SLM), rather than general-purpose models. In particular, the 'Solar Pro 1.3' version to be released this month is said to have the highest benchmark performance among domestically developed models.

CEO Kim stressed that the adoption of AI automation in the industry can lead to productivity improvement and enhanced competitiveness. He said, "Although the advancement of generative AI is fundamentally reshaping work methods across industries, many tasks are still performed by humans. Even a 1% increase in work productivity among approximately 29 million economically active people in Korea could generate an economic effect worth about 14 trillion won annually."

Since the launch of Solar last year, Upstage has secured contracts worth a total of approximately 25 billion won through building industry-specific AI models. This represents about a fivefold increase compared to 2023.

At the event, Upstage also announced its overseas expansion strategy. Recently, in collaboration with KT, Upstage built a Thai language-specialized LLM for Jasmine Technology Solutions (JTS), a Thai IT company. Following the establishment of a U.S. subsidiary last year, Upstage set up a Japanese subsidiary in March this year and is developing a Japanese language-specialized LLM in partnership with local companies.

Attending the event, Hiroyuki Matsushita, head of Upstage's Japan subsidiary, said, "The Japanese AI solution market is expected to grow to 17 trillion won by 2030. We will lead the market by strengthening locally tailored strategies based on Upstage's competitive document processing technology and small language models in the global market."

The development status of Upstage's flagship self-developed LLM 'Solar' was also introduced. CEO Kim mentioned the upcoming release of Solar 1.3, stating, "It is a model trained on a large amount of data from domestic media companies and has the highest benchmark performance among domestically developed models." Upstage plans to unveil the 'Solar Pro 1.5' version in June, which will increase the number of parameters from the current 22 billion (22B) to 33 billion, along with its first inference model.

In the subsequent Q&A session, regarding the AI copyright controversy sparked by ChatGPT's Ghibli-style art, Kim explained, "I understand that other models have used data without special permission. It feels like unfair competition. We prefer to operate legally, signing contracts with data providers and paying certain fees."

Regarding the fact that presidential candidates are releasing AI development pledges ahead of the election, he said, "I am happy that many candidates are raising AI agendas. I want to emphasize the importance of providing fair compensation and enabling the use of data for AI model training, as well as nurturing AI talent and securing graphics processing units (GPUs)."