This technical report presents K-EXAONE 2.0, an open-weight multilingual foundation model developed by LG AI Research as a step in our effort toward global frontier-scale foundation models. Rather than training from scratch, we upcycle K-EXAONE and expand its architecture, yielding a Mixture-of-Experts (MoE) model with 750B total parameters and approximately 37B activated per token---more than three times the capacity of its predecessor. K-EXAONE 2.0 supports context lengths of up to 256K tokens and expands multilingual coverage from six to ten languages. Its training pipeline combines continual pre-training, difficulty-focused mid-training, and post-training to strengthen reasoning, agentic coding, multilingual capability, and safety grounded in Korean sociocultural contexts. Across nine evaluation categories selected to reflect the conditions of practical use, K-EXAONE 2.0 improves over K-EXAONE and remains competitive with open-weight models, showing its largest gains in agentic coding and long-context understanding and its clearest strengths in long-context retrieval and safety. Released under the Apache 2.0 license, K-EXAONE 2.0 enables the wider AI ecosystem to evaluate, deploy, adapt, and build upon it, while marking the beginning---rather than the endpoint---of our challenge toward the global frontier.
Eunbi Choi, Kibong Choi, Sehyun Chun et al.· 0 citations
Developing reliable synthesis routes for complex materials remains a major bottleneck in accelerating materials discovery. This study establishes a large language model-based framework for predicting and optimizing synthesis conditions directly from the literature data. Key synthesis information, including target compounds, precursors, and processing parameters, was systematically extracted from 4407 open-access solid-state synthesis papers and organized into a structured recipe dataset. Using a retrieval-augmented generation (RAG) approach, the system first retrieves similar recipes from the corpus and then generates a new candidate recipe conditioned on those exemplars. The generated recipes were benchmarked against literature data using quantitative scoring metrics, achieving strong agreement with experimentally reported conditions. To validate the predictive capability, the framework was applied to unreported solid-state electrolyte candidates identified through first-principles screening, and multiple oxy-selenide compounds were successfully synthesized through iterative feedback between the model and experiment. The recipe generator accurately refined synthesis parameters over successive trials, demonstrating its ability to reproduce phase-pure products while minimizing trial-and-error. This approach establishes a data-driven, feedback-optimized route to accelerate synthesis design, offering a generalizable paradigm for integrating language models into experimental materials research.
Dong Won Jeon, Dong Hwi Kim, Taeyang Jeon et al.· Advances in Materials· 0 citations