LG AI Research Ships K-EXAONE 2.0 as an Apache 750B MoE
TL;DR
- LG claims end-to-end domestic development covering model architecture, distributed training, and inference infrastructure, with no foreign model distillation disclosed, a provenance claim that is material for regulated Korean procurement.
- K-EXAONE 2.0 scores 94.4 on OpenAI-MRCR long-context retrieval, beating Qwen3.5 (93.0), DeepSeek V4 Pro Max (92.9), and GLM-5.1 (71.5), the benchmarks Korean enterprise evaluators are most likely to apply.
- On a banking-specific tool-use benchmark, K-EXAONE 2.0 scores 14.2 against GLM-5.1's 11.5 and Qwen3.5's 13.4, a sector-level differentiator that matters more to financial regulators than general capability scores.
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.
What others are reporting
-
Korea Herald Read →
Reports that Exaone 4.5, a vision-language sibling already in production at Korean government ministries for public safety and drug review, gives K-EXAONE a deployment track record underpinning the procurement argument.
We have secured the capability to compete in the same weight class as global frontier models.
-
Korea Times Read →
Spotlights the banking tool-use benchmark (14.2 vs GLM-5.1's 11.5 and Qwen3.5's 13.4) as a sector-specific competitive differentiator beyond general-purpose scores, and quotes the independent-development claim directly.
Korean researchers independently completed the entire development process.
-
Korea JoongAng Daily Read →
Frames K-EXAONE 2.0 explicitly as national infrastructure under the Ministry of Science and ICT sovereign AI platform program, not a corporate R&D release, and quotes the full-stack independence claim.
Our research team independently completed every stage of development, from model architecture to large-scale distributed training.
-
The Elec Read →
Notes SK Telecom's A.X K2 shipped the same week with the identical 262K context window, placing K-EXAONE 2.0 inside an intra-Korean competition the other outlets overlook, and covers the licensing shift from proprietary to Apache 2.0.
Only 37 billion parameters are activated during inference, using a Mixture of Experts architecture to reduce computational costs.
-
Seoul Economic Daily Read →
Leads on the China comparison: K-EXAONE 2.0 at 94.4 versus GLM-5.1's 71.5 on long-context comprehension, and contextualizes Korea's state-backed AI push against China's government-led strategy.
Korean AI's global competitiveness has reached the frontier level with assessments emerging.
Originally reported by huggingface.co
Read the original article →Original headline: LG AI Research Ships K-EXAONE 2.0, a 750B/37B Korean MoE Under Apache 2.0 With 262K Context