Korea's 750B Open Model Wins 3 of the 24 Benchmarks It Published

TL;DR

LG AI Research dropped K-EXAONE 2.0 on July 31: 750 billion total parameters, 37 billion active per token, 262,144-token context, released on Hugging Face under Apache 2.0. It is the largest model Korea has ever open-sourced, and the license is a real upgrade over K-EXAONE 1.0, which shipped seven months ago under a bespoke LG agreement. Then LG did something most labs would quietly skip: it published the full scorecard. Across the 24 benchmarks on its own model card, K-EXAONE 2.0 takes the top score on three, and finishes dead last against Qwen3.5, GLM-5.1 and DeepSeek V4 Pro on seventeen.

What actually shipped

The repo name tells you the shape: K-EXAONE-2.0-750B-A37B. It is a sparse mixture-of-experts model with 256 routed experts plus one shared expert, eight experts activated per token, and an expert dimension of 2,048. The stack is 78 main layers (two dense at the head, 76 sparse) plus one multi-token-prediction layer, a hidden dimension of 6,144, 64 query heads against 8 KV heads, and a 153,600-token vocabulary. Knowledge cutoff is Q2 2025.

Attention is hybrid rather than uniformly global. The card describes one global NoPE layer and a 4,096 sliding window, then 19 repeating blocks of three 128-token sliding-window layers followed by one global layer. In plain terms: most layers only let a token talk to its immediate neighbors, and every fourth layer is the all-hands meeting where everyone gets to hear everyone. That is how you get a 262,144-token context without the KV cache eating the cluster.

Language coverage went from six to ten: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Portuguese and Polish.

The scorecard is the story

LG's press framing leads with an average of 70.1 across 24 benchmarks in nine categories, against 63.3 for K-EXAONE 1.0. That is a real generational jump and nobody should hand-wave it. But an average across 24 rows hides which rows, and the model card prints them all.

On agentic coding, the gap against the Chinese open-weight leaders is not subtle.

The same pattern repeats on SWE-bench Verified (68.2 against 76.4, 73.6 and 80.6), SciCode (37.4 against 42.0, 43.8 and 50.0), MMLU-Pro (83.5 against 89.8, 86.0 and 87.5) and Humanity's Last Exam (18.3 against 28.7, 31.0 and 37.7). The SWE-bench line is the one to sit with: K-EXAONE went 49.4 to 68.2 in one generation, a genuinely large gain, and it still lands last.

The uncomfortable one is Korean. On KMMLU-Pro, Korea's sovereign model scores 69.1 while DeepSeek V4 Pro scores 80.5. Chinese frontier models are currently better at Korean domain knowledge than the model the Korean government funded to be good at Korean.

The three rows it wins

Where K-EXAONE 2.0 does take first place, it takes it clean, and the pattern is coherent: long-context retrieval and Korean-specific safety.

The long-context number is the standout engineering result. On OpenAI-MRCR, K-EXAONE 1.0 scored 52.3. Version 2.0 scores 94.4, edging Qwen3.5 at 93.0 and DeepSeek V4 Pro at 92.9, and leaving GLM-5.1 at 71.5. Going from coin-flip retrieval to best-in-class in one generation is the kind of jump that usually comes from fixing something structural, and the hybrid attention redesign is the obvious suspect.

The safety wins are less about capability and more about whose norms are baked in. On ROK-Fortress, K-EXAONE 2.0 scores 89.5 while DeepSeek V4 Pro scores 47.6. A Korean-context safety and security benchmark is exactly the thing a Chinese-trained model has no particular reason to pass, and it is the clearest argument for why a country would spend real money on its own 750B model instead of just downloading someone else's.

The comparison LG chose to highlight

Korean press coverage led with a tool-use result: on Tau3-Bench Banking, K-EXAONE 2.0 scores 14.2, ahead of Qwen3.5 at 13.4 and GLM-5.1 at 11.5. True, and printed in the card. The same row also lists DeepSeek V4 Pro at 25.8, roughly 80 percent higher than the winner LG announced. K-EXAONE 1.0 also scored 14.2, so the number did not move at all across a generation.

None of that is fabricated, and LG deserves credit for putting the losing column in the same table as the winning one. It is just a reminder to read the row and not the press release.

Why Korea is doing this

K-EXAONE 2.0 is the second release under South Korea's Independent AI Foundation Model Project, run by the Ministry of Science and ICT. The first landed two days earlier: SK Telecom's A.X K2, a 688B-parameter MoE with 33B active, also on Hugging Face, also Apache 2.0, also with a 262,144-token context. SKT claims an average gain of 32.2 percentage points over A.X K1 across 14 benchmarks.

Lim Woo-hyung, co-head of LG AI Research, framed it directly: the point is not the parameter count, it is that Korean researchers completed every stage themselves, from architecture through data preparation to large-scale distributed training. Read against the scorecard, that is an honest positioning. This is a capability-acquisition program that shipped a usable artifact, not a claim to have caught the frontier.

Can you actually run it

Not on your desk. The model card's serving examples target two nodes of eight NVIDIA H200 GPUs, with --tp 16 for latency or --tp 8 --dp 2 for throughput, under SGLang or vLLM. The MTP layer enables self-speculative decoding, which the card puts at roughly a 3x to 5x speedup.

Reasoning mode is on by default. You pass enable_thinking=False to turn it off, and the card recommends preserve_thinking=True for agentic loops so reasoning context survives across turns. Tool calling works through an OpenAI-compatible API.

The practical read for builders: at 37B active parameters, inference economics are closer to a mid-size dense model than to a 750B one, and Apache 2.0 means you can actually deploy it commercially, fine-tune it, and redistribute it. If your workload is long-document retrieval over 200K-token inputs, or anything that has to behave correctly in a Korean regulatory or social context, this is now a serious option. If you want an agentic coding model, the card itself points you elsewhere.

Key Takeaways

• 750B total, 37B active, Apache 2.0, 262,144-token context. Korea's largest open-source model, and a license upgrade from K-EXAONE 1.0's bespoke agreement.
• Three first-place finishes out of 24 published benchmarks, and seventeen last-place finishes against Qwen3.5, GLM-5.1 and DeepSeek V4 Pro.
• OpenAI-MRCR went 52.3 to 94.4 in one generation, the strongest result on the card and now ahead of every model LG compared against.
• The other two wins are Korean safety benchmarks, where DeepSeek V4 Pro scores 47.6 on ROK-Fortress against K-EXAONE 2.0's 89.5. Sovereign alignment is the actual differentiator, not raw capability.
• Chinese models still beat it on Korean knowledge (KMMLU-Pro 69.1 vs DeepSeek V4 Pro's 80.5), which complicates the sovereignty pitch.
• Serving needs two nodes of 8x H200. Homelab-adjacent this is not, but 37B active keeps inference cost far below the parameter count's implication.

Sources: K-EXAONE-2.0-750B-A37B model card (Hugging Face), LG-AI-EXAONE/K-EXAONE (GitHub), The Korea Herald, The Korea Times, The Elec, Korea JoongAng Daily, SK Telecom Newsroom