Upstage Unveils Solar Open 100B, A Korean-Born AI Model Aiming for the Global Stage - kmjournal.net

South Korean AI startup Upstage has entered the global AI race with the release of Solar Open 100B, a large language model made public in open-weight form. According to the company, internal benchmark tests show the model delivering competitive performance against major rivals, including China’s DeepSeek R1. Solar Open 100B marks the first outcome of South Korea’s national Independent AI Foundation Model Project, led by the Ministry of Science and ICT. Since its release on December 30, the model has drawn attention not only for its performance but also for controversy. Some critics questioned whether the model relied on existing Chinese AI systems. Upstage responded by 공개ing its entire development pipeline, emphasizing that Solar Open 100B was trained from scratch without using external models. To support that claim, Upstage published technical documentation and model weights on Hugging Face, allowing researchers worldwide to inspect, test, and verify the model’s training process. Smaller Model, Competitive Results While Solar Open 100B uses far fewer parameters than DeepSeek R1, roughly 100 billion compared to 671 billion, Upstage says size did not limit performance. In internal evaluations, the model showed strong results across key language benchmarks. Upstage highlighted stable performance in Korean, English, and Japanese, with particularly strong scores in Korean-language tasks. These included cultural understanding, contextual reasoning, and knowledge grounded in Korean society. In several Korean benchmarks, Solar Open 100B reportedly outperformed OpenAI’s GPT-OSS-120B Medium, a model of similar scale. Strong Reasoning Beyond Language The company also pointed to performance in higher-level reasoning tasks. According to Upstage, Solar Open 100B demonstrated competitive results in math problem solving, complex instruction following, and agent-based tasks that require multi-step reasoning. To overcome the relative scarcity of high-quality Korean training data, Upstage relied heavily on synthetic data generation and domain-specific datasets in finance, law, and medicine. Advanced data filtering and refined training methods were applied to improve real-world usability. Efficiency Through Architecture and Training Solar Open 100B is built on a Mixture of Experts architecture, combining 129 expert models. While the total parameter count exceeds 100 billion, only a subset of parameters is activated during inference, reducing computational cost. Upstage also optimized GPU usage and applied its in-house reinforcement learning framework, SnapPO, to shorten training time and cut infrastructure expenses. A Korean AI With Global Ambitions Alongside the model release, Upstage published a detailed technical report and announced plans to open portions of its training dataset through the National Information Society Agency’s AI Hub. CEO Kim Sung-hoon described Solar Open 100B as a model deeply rooted in Korean language and culture, while remaining competitive on a global scale. He called the release an important first step toward building frontier-level AI developed entirely in Korea. With Solar Open 100B now open to external scrutiny, Upstage is positioning itself not just as a domestic AI champion, but as a serious contender in the global open AI ecosystem. by Ju-baek Shinㅣ[email protected] - Korea’s ‘National AI’ Project Faces Fresh “From Scratch” Debate as Naver Pushes Back - Upstage AI Plagiarism Controversy Comes to a Close as South Korea Calls It a Test of National AI Strength - South Korea’s Upstage Faces Plagiarism Claims Over Solar Open Model. Public Data Review Tells a Different Story - [National AI Project] Upstage Pushes Back Against Claims of Copying Chinese AI Model - [National AI Project] Upstage’s “Solar Open 100B” Faces Growing Allegations of Using a Chinese Model - Hugging Face CEO Spotlights K-AI as Korean Models Gain Global Open-Source Traction - South Korea’s AI Chiefs Call for Strict Review Amid “From Scratch” Controversy - One Year After the DeepSeek Shock: Why the Next ‘Stellar Moment’ Has Yet to Arrive