K-EXAONE 2.0 0803

K-EXAONE 2.0 0803 is a large-scale multilingual foundation model developed by LG AI Research as part of the South Korean government's Sovereign AI Foundation Model Project. Released as the second phase of the K-EXAONE initiative, it represents a significant expansion in scale and capability over its predecessor, focusing on reasoning, agentic workflows, and high-performance long-context processing. The model is released under the Apache 2.0 license, facilitating broad commercial and research adoption.

Built on a Mixture-of-Experts (MoE) architecture, K-EXAONE 2.0 contains 750 billion total parameters, with approximately 37 billion active during inference. The model was developed through an "upcycling" process, expanding the depth and width of the previous 236B K-EXAONE model rather than training from scratch. This architecture utilizes 256 experts, with 8 experts activated per token, maintaining computational efficiency while achieving frontier-level intelligence across complex tasks.

Technical Architecture and Capabilities

The model is optimized for production-ready inference through support for speculative decoding methods, including Multi-Token Prediction (MTP) and DSpark, which significantly reduce latency for long-horizon agentic tasks. It features a redesigned 150k vocabulary using SuperBPE, which improves tokenization efficiency across its supported languages. K-EXAONE 2.0 supports ten languages, including Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish, and Portuguese.

K-EXAONE 2.0 is designed with a native 256K context window (262,144 tokens), enabling the processing of extensive codebases and long-form documents. Training involved a multi-stage pipeline consisting of continual pre-training, difficulty-focused mid-training, and post-training. This approach emphasizes agentic coding and reasoning while ensuring safety grounded in Korean sociocultural contexts. Benchmarks indicate particular strength in long-context retrieval and safety compared to other leading open-weight models.