K-EXAONE 2.0 Technical Report

LG AI Research presents K-EXAONE 2.0, an open-weight Apache 2.0 multilingual foundation model built by upcycling K-EXAONE rather than training from scratch. Its sparse Mixture-of-Experts architecture expands to 750B total parameters, with 37B activated per token, while hybrid global and sliding-window attention supports contexts up to 256K tokens. The model adds depth and expert capacity, Multi-Token Prediction, and the DSpark semi-autoregressive drafter; with a draft block size of seven, DSpark reaches up to 2.57× end-to-end decoding speedup over non-speculative generation. Continual pre-training adds 8T tokens, mid-training progresses through 64K and 256K contexts, and post-training combines 350B-token supervised fine-tuning with online reinforcement learning and GROUPer preference optimization. Coverage grows from six to ten languages, including French, Italian, Polish, and Portuguese, while the Korea-Augmented Universal Taxonomy expands safety evaluation from 226 to 296 risk areas. On practical benchmarks, K-EXAONE 2.0 scores 68.2 on SWE-Bench Verified, an 18.8-point gain over K-EXAONE, and raises OpenAI-MRCR from 52.3 to 94.4, demonstrating major improvements in repository-level coding and long-context retrieval. It also reaches 99.8 on KGC-Safety, outperforming K-EXAONE, Qwen3.5, GLM-5.1, and DeepSeek V4 Pro on that Korean-context safety benchmark, although it does not lead every general reasoning or coding comparison.