LG AI Research Ships K-EXAONE 2.0 as an Apache 750B MoE

The interesting part of LG AI Research's K-EXAONE 2.0 release on Hugging Face is not the parameter count but the licence. A 750-billion-parameter mixture-of-experts model with 37 billion active parameters, shipped under Apache 2.0 by a Korean AI lab, is a different shape of open-weight release than what US and Chinese labs have been producing this year.

The architecture, as documented on the model card, uses 256 experts plus one shared expert with 8 activated per token, a 262,144-token context window, and support for ten languages including Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish and Portuguese. LG reports 83.5 on MMLU-Pro, 92.3 on AIME 2026, 68.2 on SWE-Bench Verified and 94.4 on OpenAI-MRCR for long context. The card also describes two speculative decoding methods, MTP and DSpark, which LG says can accelerate generation by roughly 3 to 5x for long-horizon agentic workloads.

Why this matters if you are not training models yourself: the pool of genuinely capable, permissively licensed base models has been narrower than the release cadence suggests. Most open weights arrive with source-availability caveats, non-commercial clauses, or the quiet understanding that they trail the frontier by a step. An Apache-2.0 model from a well-resourced non-US, non-Chinese lab, with credible multilingual coverage and a quarter-million-token context, gives enterprise teams in Europe and Asia a new default candidate to evaluate before reaching for Llama or Qwen.

The honest caveat is that these numbers are LG's own, on benchmarks LG chose to publish, with the knowledge cutoff listed as Q2 2025. What the model card does not give you is any independent evaluation on identical harnesses, a realistic cost of serving the 750B/37B configuration in production, or safety results beyond the single internal score reported.

If the scores hold up outside LG's own harness, the story worth watching is the geographic diversification of who ships credible open weights, not any one benchmark line.