LG's K-EXAONE 2.0 puts 750B open weights on Hugging Face

LG AI Research released K-EXAONE 2.0 on Hugging Face on July 31, the second model from South Korea's Sovereign AI Foundation Model Project. The official announcement calls it Korea's largest AI foundation model and says the lab completed the training and inference infrastructure needed to build it. 1

One number needs precision: the model is 750B parameters, with 37B active parameters per inference path, not 75B. It is a mixture-of-experts model, expanded from the first K-EXAONE's 236B total parameters. The model card lists a 262,144-token context window, a 2025 Q2 knowledge cutoff, and support for Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish, and Portuguese. 2

LG reports an average score of 70.1 across 24 benchmarks in nine categories, up from 63.3 for the first model. The largest reported gains are in coding and agentic coding, where three headline evaluations improved by about 30%. Its clearest lead is in long-context retrieval: the model card reports 94.4 on OpenAI-MRCR versus 71.5 for GLM-5.1, and 89.6 on Ko-LongBench versus 83.6. It also scores 14.2 on τ3-Bench Banking, ahead of GLM-5.1 at 11.5 and Qwen3.5 at 13.4. These are LG's published evaluations, not independent results. 1 2

The table also argues against reading the release as a universal win. K-EXAONE 2.0 scores 68.2 on SWE-Bench Verified, below Qwen3.5's 76.4, and 92.3 on AIME 2026, below DeepSeek V4 Pro Max's 95.2. Its profile is stronger in long-context and selected agentic tasks than across every general benchmark. 2

LG is releasing the model under Apache 2.0, which permits commercial use, and the Hugging Face collection lists BF16, FP8, NVFP4, and DSpark variants. The access story is still demanding: LG's serving examples use two nodes with eight NVIDIA H200 GPUs each. The 37B active-parameter figure reduces the experts used for each token; it does not make a 750B checkpoint small to store or distribute. The model card also flags a current B200 workaround and says DSpark serving is not yet supported in vLLM. 1 2 3

That makes the launch meaningful as an open-weight and sovereign-AI move, but not yet a casual local download. Developers gain a commercially usable frontier-scale option; most teams will still need a hosted provider or serious multi-GPU infrastructure before they can test it. LG says a public evaluation platform and a service for trying K-EXAONE 2.0 are in preparation. 1