Upstage has unveiled 'Solar Open2,' a large language model (LLM) aimed at its proprietary artificial intelligence (AI) foundation model. This model focuses on creating AI agents capable of performing real work tasks, such as document writing, Excel analysis, presentation creation, and coding, as part of its strategy to penetrate the enterprise AI market.
On July 22, Upstage held a developer conference, 'Solar Open Weight Day,' where it introduced Solar Open2. The model was made available for developers through an open weight distribution via Hugging Face.
Sung-hoon Kim, CEO of Upstage, stated at the event, “These days, it’s not enough to just have good conversations. If you can’t perform tasks when asked, you’re treated like a fool. That’s why we focused on developing the capabilities of AI agents.”
A key differentiator of this model is the custom simulation environment called 'Officeverse,' designed to train the AI in real work settings. Officeverse creates virtual scenarios for tasks like document writing, Excel work, and PPT creation, allowing the AI to be evaluated on its accuracy in performing these tasks.
Solar Open2 features a total of 250 billion parameters. It employs a Mixture of Experts (MoE) structure, activating only 15 billion parameters during the actual inference process. This design maintains high performance while reducing GPU usage and inference costs.
Upstage emphasized that Solar Open2 can handle contexts of up to 1 million tokens, making it suitable for analyzing large documents such as contracts and reports in one go. The model has enhanced features necessary for implementing AI agents, including document creation, external tool invocation, and executing complex workflows.
Performance competitiveness was also highlighted. According to data released by Upstage, Solar Open2 scored 92.4 on the code generation performance benchmark 'LiveCodeBench,' surpassing the scores of comparable models like DeepSpeed V4-Flash (92.3) and Mistral Medium 3.5 (84.9).
With the launch of Solar Open2, Upstage plans to intensify its efforts in the enterprise AI agent market. The company aims to expand its presence in the domestic AI ecosystem, leveraging the large model development capabilities gained through participation in the government’s proprietary AI foundation model project.
When asked about monetization strategies following the free release, Kim emphasized 'ecosystem expansion.' He remarked, “Right now, the most important thing in Korea is the virtuous cycle of the AI ecosystem. Our goal is for Solar Open2 to be utilized in real work environments and to be recognized as a good model, thereby creating a larger ecosystem.”
He added, “Through support from the government’s proprietary AI foundation model project, we were able to develop a model of 250 billion parameters following our previous 100 billion model. Our focus is not just on creating large models but on developing practical AIs that can be used in real-world applications.”
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.