Analysis
LG AI Research released K-EXAONE 2.0 on Hugging Face on July 31, a 750-billion-parameter frontier model that more than triples the 236 billion parameters of the original EXAONE and is now South Korea's largest publicly released AI model. The model uses a hybrid-attention mixture-of-experts architecture, activating roughly 37 billion of its 750 billion total parameters per token, and LG released it under an unrestricted Apache 2.0 license.
The licensing choice matters as much as the parameter count. Apache 2.0 permits commercial use without the usage restrictions frontier labs like Meta attach to Llama, putting LG in the same fully-open camp as DeepSeek and Moonshot AI rather than the closed-API model OpenAI, Anthropic and Google run. LG's first EXAONE model drew a fraction of the attention Chinese open-weight releases have gotten this year; a 750-billion-parameter flagship, unrestricted, is a deliberate attempt to compete on the same playing field.
“The licensing choice matters as much as the parameter count.”
On benchmarks, K-EXAONE 2.0 posted a 70.1-point average across 24 evaluation categories, more than a 10% jump over its predecessor, with the sharpest gains -- roughly 30% -- in coding and agentic-coding tasks specifically. The model also ships with a 262,144-token context window and support for 10 languages including Korean, English, Japanese and Chinese, positioning it as a direct entrant in the same open-weight race DeepSeek's V4 Flash and Moonshot's Kimi models are already running.
The release lands the same week DeepSeek pushed out V4 Flash 0731 at unchanged $0.14/$0.28-per-million-token pricing, underscoring how crowded and price-competitive the open-weight tier has become. For US labs, the pattern is now familiar: every few weeks a well-funded Asian lab ships a frontier-scale open model that undercuts closed-API pricing entirely by being free to self-host, and enterprises building on top of frontier models increasingly have a credible non-US, non-API alternative.
What to watch is adoption, not benchmarks -- open-weight releases only matter commercially if developers actually build on them instead of just downloading and testing them. LG's own enterprise relationships across South Korea's chaebol ecosystem give K-EXAONE 2.0 a plausible enterprise on-ramp that pure research labs lack, which is the real test of whether this becomes infrastructure or a benchmark headline.
