TOTORONG/Solon_Athens_v5_STEM-31B
Solon_Athens_v5_STEM-31B is a 31 billion parameter large language model developed by Nextnine, based on Google's Gemma4 architecture with Hybrid Attention. It is fine-tuned for specialized performance in STEM (Science, Technology, Engineering, Mathematics) and Korean civil law domains, utilizing a Cascading Learning technique to minimize catastrophic forgetting. This model excels at complex reasoning and problem-solving within its specialized fields, maintaining general knowledge while acquiring new expert domain understanding.
Loading preview...
Model Overview
Solon_Athens_v5_STEM-31B is a 31 billion parameter large language model developed by Nextnine, based on Google's google/gemma-4-31B-it. It is specifically fine-tuned for STEM (Science, Technology, Engineering, Mathematics) and Korean civil law domains.
Key Differentiator: Cascading Learning
A core innovation of this model is Nextnine's proprietary Cascading Learning technique. This method enables the model to sequentially learn multiple specialized domains without significant catastrophic forgetting of previously acquired general knowledge or other domain-specific expertise. This ensures the model retains its broad capabilities while effectively integrating new, specialized knowledge.
Key Capabilities
- STEM Reasoning: Excels in problem-solving and question-answering related to mathematics, science, and technology.
- Korean Legal Understanding: Proficient in analyzing, summarizing, and reviewing documents within the civil law domain.
- Chain-of-Thought (CoT) Research: Designed to evaluate and demonstrate step-by-step reasoning processes.
- Continual Learning: Showcases the effectiveness of the Cascading Learning approach in maintaining knowledge across sequential learning tasks.
Performance Highlights
As of June 12, 2026, Solon_Athens_v5_STEM-31B achieved 2nd place overall on the K-AI Leaderboard, with notable scores including 0.702 on KMMLU-Pro (Korean multi-domain reasoning) and 0.853 on CLIcK (Korean conversation and knowledge reasoning).
Intended Use
This model is primarily intended for research and evaluation purposes. It is not recommended for high-stakes applications such as legal advice, medical diagnosis, or financial decisions. Performance outside of Korean and the specialized STEM/legal domains may be limited to the base model's capabilities.