Korea’s representative artificial intelligence (AI) startup upstage announced that it can launch a brand new AI model in succession and construct a ‘domain -specific B2B AI Agent’. The secret is to mix the strength of optical character recognition (OCR) and the representative model ‘Solar’, which is targeting the company document market with a robust multimodal function.
Upstage (CEO Kim Seong -hoon) held a gathering on the Yeouido Conrad Hotel on the sixteenth to announce the present status of business and technology development and future plans.
Kim Seong -hoon said he’ll launch a lot of models this 12 months. First, we released the Solar Pro 1.3 version.
He stressed that benchmark performance is the very best among the many upstage models. It is best for specific answers and systemization than the prevailing version.
In June, the corporate plans to unveil the ‘Solar Pro 1.5’ version, which has expanded to 33 billion (33B) in the prevailing 22 billion (22b) parameters, and the primary reason for implementing the ‘COT’.
Solar 1.5 will deal with convenient usability, comparable to ‘fluent sentence speech’ moderately than 1.3 and supports the model concept protocol (MCP).
Kim emphasized these models that “it just isn’t a general model, but has a novel technical competitiveness in the world of specialized small language models (SLM) by industry.”
Subsequently, it can expand right into a multimodal model (LMM) that integrates OCR and enormous language model (LLM). In June, the OCR model ‘DOCVLM’, which mixes the OCR model ‘Document Pas (DP)’ and the CEO LLM Solar, which have a specialty in complex layout recognition, might be unveiled in June.
This model is nice at various documents based on information summaries, questions and reports. In consequence of the test, he claimed that it was more accurate than Meta’s Lama 4 Scouts and Google’s ‘Geminai 2.5 Pro’.
“In case you process corporate document data using LMM, you’ll be able to quickly handle text and layouts of 100 to 1000 pages,” he said.
Particularly, “the B2B AI model has a transparent purpose and demand, unlike general -purpose models,” he said.
Inquiries of actual DPs are increasing rapidly. In May, it can be released in the shape of ‘Work Space’ and can enable you experience performance easily. For instance, for those who put 100 documents, you’ll be able to organize every time by time zone.
He said he can also be specializing in data. “The reasoning model that requires quality data is increasing its performance mainly on mathematics and coding data,” said Lee Boo -seok, an upstage CTO. It also said it’s partnering with firms to gather legal and ethical data.
Overseas businesses are also actively developing, comparable to the localization of solutions with Japan and US subsidiaries. Particularly, there aren’t any DP competitors in Japan, and so they are actively targeting the document market in Japan.
As well as, we’re discussing collaboration with the Middle East, and plans to enter Southeast Asia in earnest by developing Thai LLM development, which recently collaborated with KT.
Kim Sung -hoon, CEO of Upstay, said, “As we deal with the event of model, we may have numerous infrastructure funds comparable to GPUs.”
By Jang Se -min, reporter semim99@aitimes.com
