Upstage has announced significant performance improvements for its upcoming open-source large language model (LLM), "SOLAR Open2," as it prepares for the second-phase evaluation of the South Korean government's "Dokpamo" (독파모, Independent AI Foundation Model) project.

According to Kim Sung-hoon, CEO of Upstage, "SOLAR Open2" demonstrates a 2 to 3 times performance enhancement compared to its predecessor, "SOLAR Open1." This finding is based on ablation studies, a critical engineering process that involves systematically removing or altering components of a model to assess their individual contribution to overall performance and validate design choices prior to final release.

The development of SOLAR Open2 is heavily leveraging advanced training methodologies. Upstage is actively recruiting for roles in post-training and visual language models (VLM), with specific emphasis on experience in post-training and Reinforcement Learning from Validated Rewards (RLVR). Post-training is the process of further fine-tuning a pre-trained model to understand user commands and become applicable in real-world services. This typically involves techniques like Supervised Fine-Tuning (SFT) and other forms of reinforcement learning. RLVR is a specialized reinforcement learning technique that boosts inference performance by rewarding the model for correctly solving problems with clear, verifiable answers, such as in mathematics or coding, as notably utilized by models like DeepSeek.

SOLAR Open2 is expected to continue Upstage's "lightweight high-performance strategy," which was exemplified by "SOLAR Open 100B" in the first phase of the Dokpamo project. "SOLAR Open 100B," with its 100 billion parameters ($10^{11}$ parameters), achieved performance comparable to much larger models, emphasizing efficiency. It secured a perfect score (10 out of 10) in the global individual benchmark section during the initial Dokpamo evaluation, a feat matched only by LG AI Research.

The Dokpamo project, initiated by the Ministry of Science and ICT in August of the previous year, aims to mitigate technological, cultural, and economic dependence on global AI models. Initially, five elite teams competed, with LG AI Research, SK Telecom, and Upstage advancing to the second phase after the January evaluation. Motif Technology has since joined the project through a subsequent open call. These models, including Upstage's SOLAR Open 100B, LG AI Research's K-Exaone-236B, and SK Telecom's A.X K1, have gained international recognition, being listed among Epoch AI's "Notable AI Models" and prominently featured on Hugging Face's trending models page, with Hugging Face CEO Clement Delangue personally promoting them.