K-EXAONE 2.0: LG AI Research Releases a 750B Open-Weights MoE Model Under Apache 2.0

On July 31, 2026, LG AI Research published the weights of its large language model K-EXAONE 2.0 on Hugging Face¹². The license is Apache License 2.0, so commercial use is permitted under its terms¹.

The model has 750B total parameters with 37B active per token, in a Mixture of Experts (MoE) architecture¹. It arrives in a month that also saw Moonshot AI release the full weights of Kimi K3 and Alibaba announce that Qwen 3.8’s weights would follow — now with a Korean entry alongside them.

What the model card specifies

The published model card is unusually specific about the configuration¹.

Item	Value
Total / active parameters	750B / 37B
Layers	78 (2 leading Dense + 76 Sparse) plus 1 MTP layer
Experts	1 shared, 256 total, 8 activated (expert dimension 2,048)
Hidden dimension	6,144
Vocabulary size	153,600
Context length	262,144 tokens
Knowledge cutoff	Q2 2025

Attention is arranged as one Global (NoPE) layer and one 4,096 SWA layer, followed by nineteen repetitions of a [3 × 128 SWA + 1 × Global] block, with 64 Q-heads, 8 KV-heads, and a head dimension of 128¹. MoE activates only part of the total parameters for each token, which is what keeps inference compute at the 37B level despite the 750B total.

The supported languages are Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish, and Portuguese — ten in all¹. Press reports say French, Italian, Portuguese, and Polish are new relative to the previous generation³. The model card also states that it supports speculative decoding, with a speedup of approximately 3–5×¹.

There is more than one artifact: alongside the standard release, FP8, NVFP4, and DSpark variants are published in the same Hugging Face org, four in total². Having quantized versions on day one matters for anyone deciding whether this can actually run on the GPUs they have.

The published scores, and how to read them

The model card lists benchmark results¹: MMLU-Pro at 83.5, GPQA-Diamond at 82.2, SWE Bench Verified at 68.2, long-context OpenAI-MRCR at 94.4, Korean-language KMMLU-Pro at 69.1, multilingual MMMLU at 86.6, and KGC-Safety at 99.8¹.

Press reports put the average across 24 benchmarks in nine categories at 70.1, an improvement of more than 10% over the previous generation³. K-EXAONE 1.0 had 236 billion parameters, so 2.0 is more than three times its size³. On long-context comparisons, K-EXAONE 2.0 reportedly scored 94.4 on OpenAI-MRCR against 71.5 for China’s GLM-5.1, and 89.6 against 83.6 on the Korean-language Ko-LongBench³.

These are all published figures, not numbers independently reproduced by third parties. Given what a benchmark score does and does not measure, checking whether the comparison conditions match your own workload is faster than reasoning about the table. That is precisely what open weights make possible.

”The same weight class as global frontier models”

Lim Woo-hyung, co-head of LG AI Research, said the organization has “secured the capability to compete in the same weight class as global frontier” models, according to press reports³. The models named as comparisons are China’s GLM-5.1, Qwen3.5, and DeepSeek V4 Pro Max³.

The Korea Herald describes K-EXAONE 2.0 as the second model developed under a Korean state-led program to build domestic foundation models³. Seoul Economic Daily, covering the same release, does not explicitly characterize it as government-funded or as a formal sovereign AI project⁴. Because the reporting does not line up, the relationship to that program cannot be stated definitively. The company is also reported to be planning an industry-specialized model next week⁴.

What is concrete is that one more 750B-class set of weights is now available under a license that permits commercial use — a practical change for anyone weighing on-premises or private-cloud deployment. The Apache 2.0 terms and which of the four variants fits a given environment can both be checked on Hugging Face right now¹².