Seven Hundred Fifty Billion Parameters, Just Handed Over

On July 31, a repository quietly appeared on Hugging Face: LGAI-EXAONE/K-EXAONE-2.0-750B-A37B. The name tells you almost everything. Seven hundred fifty billion total parameters. Thirty-seven billion of them actually firing for any given token. It is the largest AI foundation model ever released in Korea, and LG AI Research put it out under Apache 2.0. Download it, dissect it, fine-tune it, build a business on top of it, sell that business — LG gets nothing and asks for nothing. That is what happened on Friday.

Here's the deal: you need to understand how strange that is before the size means anything. Only months earlier, LG's first-generation K-EXAONE shipped under something called the "K-EXAONE AI Model License Agreement." A house license is always a fence. It says: use this inside the boundaries we drew. And then the second-generation model, more than three times bigger, arrived with the fence torn down entirely. Apache 2.0 is about as permissive as open-source licensing gets. It includes an explicit patent grant, which is the specific clause enterprise legal departments care about most. Going from a bespoke license to Apache 2.0 is not a minor line item in a release note. It is a strategic reversal.

So why now? Look at the calendar, not the model card. K-EXAONE 2.0 is the second-round deliverable for the Independent AI Foundation Model Project run by Korea's Ministry of Science and ICT — the national program Koreans shorthand as "dokpamo." The second evaluation round is scheduled for early August. One of the surviving teams gets cut. Which means this release is not a pure technical event. It is an exam answer sheet, submitted days before grading, published in the open so that everyone can watch it get scored. That framing explains details that otherwise look odd, including which benchmarks got emphasized and which quietly did not.

Lim Woo-hyung, co-head of LG AI Research, framed it exactly that way in his public statement: "The significance of K-EXAONE 2.0 is not simply that it is a large-parameter model, but that Korean researchers independently completed the entire development process." In other reporting he expanded that to cover the full chain — architecture design, data preparation, large-scale distributed training, inference infrastructure. Notice what he is arguing. He is not primarily arguing that the model is good. He is arguing that Koreans built all of it themselves. In a national sovereignty program where "independence" is a scored criterion, that is the sentence that counts.

And LG was not alone in the window. Two days earlier, on July 29, SK Telecom posted A.X K2 — 688 billion parameters, also Apache 2.0, also on Hugging Face. Eight days before that, on July 23, Upstage released Solar Open 2 at 250 billion parameters. Three large Korean open-weight models in a nine-day stretch. That is not a coincidence of research calendars. That is four teams sprinting toward the same August checkpoint, and only some of them will still be standing afterward.

Who's Actually In This Fight

Start with LG AI Research itself. Founded in 2020 as the LG Group's dedicated AI institute, it has run the EXAONE line longer than almost anyone else in Korea has run a serious LLM program. EXAONE 3.0 in 2024 was its first real open-source entry. EXAONE 4.0 in 2025 introduced the hybrid attention approach that shows up again here. Then came the first-round K-EXAONE at the end of last year: 236 billion total parameters, 23 billion active, a Mixture-of-Experts design. LG says that model ranked seventh globally and first in Korea among open-weight models on the Artificial Analysis Intelligence Index at the time, and briefly climbed to second on Hugging Face's trending list. For a Korean team entering a field dominated by American and Chinese labs, that was a genuinely strong debut, and it is why LG went into the second round as the team to beat.

Then there is the Ministry of Science and ICT, which designed the game everyone is playing. The stated goal is blunt: reduce Korea's technological, economic, and security dependence on foreign models like GPT and Gemini. The mechanism is a survival tournament. Pick elite teams, hand them national GPU capacity, evaluate them in stages, and eliminate the weakest each round. The compute subsidy is not symbolic — Motif Technologies reportedly built a preview model in roughly five months using around 700 government-supplied GPUs. For companies that could never justify that capital expenditure alone, the state is effectively underwriting the training run.

The tournament has already claimed casualties. Five elite teams started. In the first evaluation, the Naver Cloud consortium and NC AI were cut, with the assessment concluding they did not clear the independence bar. Naver getting eliminated was a shock at the time, because Naver had effectively pioneered Korean-language LLM work with HyperCLOVA X. The ministry then reopened applications to backfill, bringing in Motif Technologies and Trillion Labs. The current field is four: LG AI Research, SK Telecom, Upstage, and Motif Technologies. In early August one more goes home, and the final two teams are expected to be confirmed next February.

That structure is the missing context in most coverage. If you only see "Korean companies released three huge models in a week," it reads as a burst of national enthusiasm. It is not. It is four teams pushing their strongest possible artifact onto a public platform right before a scored review, because a Hugging Face release is itself a form of evidence. You cannot argue with weights that anyone can download. Publishing under Apache 2.0 adds a second layer of evidence: it demonstrates confidence, and it removes any suspicion that a nationally funded model is being quietly enclosed.

Finally, look at who LG chose as comparison models on the model card, because that choice is revealing. Qwen3.5. GLM-5.1. DeepSeek V4 Pro. Not a single American frontier model appears in the table. That is an accurate read of the terrain rather than an evasion. The absolute top of the capability curve sits behind closed APIs in the United States, while the open-weight frontier — the models you can actually download — has been overwhelmingly Chinese for the past couple of years. Korea is not competing for the closed crown. It is competing for shelf space in the open-weight tier, and that means the opponents are Alibaba, Zhipu, and DeepSeek.

What's Actually Inside

The architecture is a hybrid attention Mixture-of-Experts, and the layer layout is unusual enough to be worth spelling out. There are 78 layers total: two dense layers at the head, 76 sparse main layers after them, and one multi-token-prediction layer at the end. The MoE block holds 256 experts with 8 activated per token, plus one shared expert that stays on for every token regardless of routing. The attention schedule starts with one global layer using no positional encoding and one sliding-window layer at 4096, then repeats a block of three 128-width sliding-window layers followed by one global layer, nineteen times over. That pattern — dense local attention, sparse global attention — is the standard modern recipe for making very long contexts affordable, and here it is deployed at scale.

The rest of the shape: 64 query heads against 8 key-value heads, head dimension 128, hidden dimension 6,144, intermediate dimension 18,432, expert dimension 2,048, vocabulary of 153,600 tokens. Context window is 262,144 tokens. The knowledge cutoff is Q2 2025, which is a detail worth internalizing before you test it — nothing from the first half of 2026 exists inside the weights, so any question about recent events will get a confident fabrication unless you wire in retrieval or tools. As for how it was built, the model card says K-EXAONE 2.0 was "scaled to more than three times the size of its predecessor through upcycling, followed by continual pretraining, difficulty-focused mid-training, and post-training." Read that carefully. This is not a from-scratch 750B training run. It is the 236B model grown into a larger sparse architecture and then trained further. That is a legitimate and increasingly common technique, and it is also considerably cheaper than starting over.

Now the scores. Across 24 benchmarks spanning nine categories, K-EXAONE 2.0 averages 70.1, up from 63.3 for the first-generation model — better than a 10 percent relative gain. Coding and agentic coding improved by roughly 30 percent on average. But averages hide the interesting part, and the interesting part is the distribution. Here is the table from LG's own model card. Every number below is self-reported by LG. There is no third-party audit of these results, and I want that stated before you read a single row.

Category Benchmark K-EXAONE 2.0 K-EXAONE 1.0 Qwen3.5 GLM-5.1 DeepSeek V4 Pro
World knowledge MMLU-Pro 83.5 83.8 89.8 86.0 87.5
World knowledge GPQA-Diamond 82.2 79.1 88.4 86.2 90.1
World knowledge Humanity's Last Exam 18.3 13.6 28.7 31.0 37.7
Math AIME 2026 92.3 92.2 91.3 95.3 95.2
Math HMMT Feb 2026 78.4 80.7 84.6 82.6 95.2
Math IMO Answer 78.6 76.3 80.9 83.8 89.8
Coding SWE-Bench Verified 68.2 49.4 76.4 73.6 80.6
Coding Terminal-Bench 2.1 43.8 30.3 51.3 61.8 64.0
Coding SciCode 37.4 35.6 42.0 43.8 50.0
Tool use τ³-Banking 14.2 14.2 13.4 11.5 25.8
Tool use Claw-Eval 77.7 70.3 79.7 84.4 82.7
Instruction following IFEval 92.4 89.7 92.6 93.9 94.0
Instruction following IFBench 72.6 67.3 76.5 76.3 76.5
Long context OpenAI-MRCR 94.4 52.3 93.0 71.5 92.9
Long context AA-LCR 56.2 53.5 65.7 62.3 66.3
Long context Ko-LongBench 89.6 86.8 91.3 83.6 91.4
Korean KMMLU-Pro 69.1 67.3 77.4 75.8 80.5
Multilingual GlobalMMLU-Lite 86.6 86.9 92.1 90.7 92.0
Safety KGC-Safety 99.8 96.1 92.0 69.3 82.8
Safety ROK-Fortress 89.5 60.9 86.1 73.2 47.6

Turns out the table tells a much more specific story than the press release does. The OpenAI-MRCR result LG led with is real and it is dramatic: 94.4, up from 52.3 in the first-generation model, ahead of Qwen3.5 at 93.0 and DeepSeek V4 Pro at 92.9, and miles past GLM-5.1 at 71.5. Nearly doubling a long-context retrieval score in one generation is not something you get by accident; that is a targeted engineering result, and the attention schedule described above is presumably why. Safety is the other standout. KGC-Safety at 99.8 and ROK-Fortress at 89.5 are commanding numbers, especially against DeepSeek V4 Pro's 47.6 on ROK-Fortress. These are Korea-specific safety and compliance evaluations, and LG clearly tuned hard for them.

Then look at the rest and the picture cools considerably. MMLU-Pro went from 83.8 to 83.5 — a model that tripled in parameters got very slightly worse on a headline knowledge benchmark. HMMT dropped from 80.7 to 78.4. Humanity's Last Exam sits at 18.3 while DeepSeek V4 Pro posts 37.7, more than double. SWE-Bench Verified jumped impressively from 49.4 to 68.2, which is a real generational leap, but it still trails Qwen3.5 at 76.4 and DeepSeek at 80.6. Terminal-Bench 2.1 at 43.8 is far behind GLM-5.1's 61.8. Agentic tool use on τ³-Banking is 14.2, literally unchanged from the previous generation, against DeepSeek's 25.8. The honest summary is that K-EXAONE 2.0 is not a model that closed the frontier gap across the board. It is a model that concentrated its gains in long-context comprehension and Korea-specific safety while holding roughly steady elsewhere. That is a defensible strategy for a sovereign model. It is just not the same claim as the headlines imply.

And now the part that determines whether any of this matters in practice. The serving example in the model card calls for two nodes of eight NVIDIA H200 GPUs each — sixteen GPUs. You cannot use stock inference stacks either; the documentation points to custom forks of both SGLang and vLLM maintained by LG engineers. Tensor parallelism is tp=16 for the low-latency configuration, or tp=8 with dp=2 for throughput. Speculative decoding via EAGLE/MTP weights is reported to deliver a 3–5x speedup. Recommended sampling is temperature 1.0 and top_p 0.95, with reasoning mode enabled by default. One more thing worth flagging: the citation block on the model card still reads arXiv:XXXX.XXXXX. The technical report has not shipped. The model card explicitly says detailed evaluation configurations live in that report — which means, as of right now, there is no public way to verify how any of the table above was measured.

Who Wins What

LG's payoff comes in three layers. The immediate one is the exam. Walking into an early-August evaluation holding the largest Korean model ever built, the broadest language coverage of any domestic foundation model, and the strongest safety numbers in the comparison set gives evaluators very little to push back on. The second layer is political cover. A model built partly with public money and public GPUs, locked behind a proprietary house license, invites an obvious and repeated question about who owns nationally funded research. Apache 2.0 makes that question disappear in one move. The third layer is ecosystem compounding. Permissively licensed weights generate fine-tunes, quantizations, third-party evaluations, and integration guides without LG spending a won on any of it, and every one of those artifacts becomes a citation for LG.

The Ministry of Science and ICT gets legitimacy, which it has needed. The dokpamo program has drawn skepticism since inception about whether a government should be funding frontier model training at all, and the first-round elimination of Naver triggered a separate round of criticism about the evaluation criteria themselves. Now, in a nine-day window, the program has produced a 750B model, a 688B model, and a 250B model, all published openly, most under Apache 2.0. As proof that the machine produces output, that is hard to argue with. Whether the output gets used is a different question and a harder one, but for the immediate purpose of defending a budget line, shipping is most of the argument.

Korean enterprises get something genuinely valuable, at least on paper. A top-tier open-weight model, tuned for Korean, with no licensing friction and no per-token bill, that can run entirely inside your own data center. For banks, hospitals, and government agencies operating under data residency rules that forbid sending anything to a foreign API, the value of a capable on-premises model is not incremental — it is the difference between having an option and not having one. Until now, organizations in that position were choosing between much smaller domestic models and simply not deploying. A 750B-class option changes the shape of that decision. The catch, again, is sixteen H200s.

Researchers get the most unambiguous gift here. The weights of a 750B hybrid-attention MoE are now openly available with no usage restrictions. You can run ablations on the attention schedule and measure exactly how much that global-plus-sliding-window arrangement contributes to the OpenAI-MRCR jump. You can study what upcycling a 236B model into a 750B sparse architecture actually does to capability distribution, which is a recipe frontier labs rarely publish in any detail. The constraint is the missing technical report — until it lands, the data mixture, the mid-training curriculum, and the evaluation harness all remain opaque, and much of the most useful analysis has to wait.

There is also a quieter beneficiary: Korea's domestic AI chip industry. When the first-generation K-EXAONE shipped, FuriosaAI published NVFP4+GPTQ quantized weights for it. If something similar happens for the 750B model, the story stops being "Korea has a big model" and becomes "Korea has a big model that runs on Korean silicon," which is a materially different industrial narrative. Nothing has been announced on that front yet, and it is one of the more consequential things to watch over the next few months.

We've Seen This Movie Before — Both Endings

Start with the successes, because they share a pattern. Meta's Llama series is the canonical case: the moment Llama 2 allowed commercial use, the fine-tuning ecosystem exploded, and for a stretch "open-source LLM" and "Llama derivative" were nearly synonymous. Alibaba's Qwen took the lesson further by shipping a dense ladder of sizes — tiny models you can run on a laptop all the way up to flagship scale — under Apache 2.0, and it now anchors more derivative models on Hugging Face than any other family. DeepSeek won differently, by making efficiency itself the product and publishing enough methodology to reset the industry's assumptions about training cost. Different strategies, one common thread: none of them just dropped weights. They shipped a size ladder, a toolchain, documentation, and a predictable release cadence.

The failures are equally instructive. TII's Falcon 180B from the UAE launched as the largest open model available and was marketed heavily as a sovereign achievement. The ecosystem barely materialized. The reasons were mundane rather than mysterious: it was too large for most people to serve, its license carried conditions that made legal review annoying, and follow-up support was thin. BigScience's BLOOM at 176B was an admirable scientific effort with excellent multilingual coverage and very little real-world adoption. Meta's own OPT-175B was released under a research-only license that capped its influence before it started. Three projects, three different flavors of the same outcome: enormous models that made headlines and then sat mostly unused.

So where does K-EXAONE 2.0 land on that spectrum? On licensing, it is at the good end — Apache 2.0 eliminates the friction that hurt Falcon and killed OPT's momentum. It has real differentiation in Korean-language and Korea-specific safety performance, plus a genuinely strong long-context capability at 262K tokens, which is a practical rather than cosmetic advantage. But the warning signs are visible too. The serving footprint is enormous. The inference path requires custom forks rather than upstream releases. The technical report is missing. And most importantly, there is no smaller sibling. Llama and Qwen won largely on the strength of 7B and 32B checkpoints that individual developers could actually run, and K-EXAONE 2.0 currently has no equivalent.

The first-generation model is the useful reference point here. At 236B/23B, K-EXAONE 1.0 was within reach of community tinkering — Q4_K_M GGUF quantizations appeared, and FuriosaAI produced hardware-specific weights. Someone with a serious workstation or a modest cloud budget could engage with it. At 750B, that door mostly closes. The decision to scale up is fully consistent with the national program's stated objective of reaching frontier weight class, and I do not think it was the wrong call for the exam. But "matching the frontier in scale" and "building an ecosystem" pull in different directions, and that tension is the single most important unresolved thing about this release.

One more historical note that cuts the other way: sovereign models sometimes succeed on distribution rather than downloads. If the Korean government mandates or strongly prefers domestic models for public-sector AI procurement, adoption can be manufactured regardless of whether hobbyists ever touch the weights. LG already has evidence in that direction — its EXAONE 4.5 vision-language model was selected to analyze the more than 39,000 daily public safety reports handled by the Ministry of the Interior and Safety, and to support drug application reviews at the Ministry of Food and Drug Safety. That channel does not require anyone to own sixteen H200s. It requires the government to keep choosing LG.

How the Competition Punches Back

The most immediate counterplay is inside the same tournament. SK Telecom's A.X K2, released July 29, runs 688 billion total parameters with 33 billion active, a 262,144-token context, and — critically — the same Apache 2.0 license. SKT says A.X K2 improved an average of 32.2 percentage points over its predecessor A.X K1 (519B) across 14 domestic and international benchmarks, with roughly 83.9 percentage points of improvement on long-context and agentic evaluations. Those improvement figures depend entirely on the baseline they are measured against, so I would not put them next to LG's absolute scores and call it a comparison. The point is that scale, context length, and licensing are now effectively matched between the two largest teams. Whatever separates them in August will be something other than headline specs.

Upstage attacked from a completely different angle, and it is the more interesting strategic bet. Solar Open 2, released July 23, is 250 billion total with only 15 billion active — far smaller. In exchange it offers up to a one-million-token context, and Upstage says a quantized version runs on just two H200 GPUs. It also arrived with a concrete deployment story attached: integration into the Daum portal. That is a deliberate contrast in positioning. LG is arguing weight class; Upstage is arguing deployability. If the August evaluation weights raw capability, LG is comfortable. If it weights practical deployment and cost of operation, sixteen H200s starts looking like a liability rather than an achievement. Motif Technologies, meanwhile, brings speed as its differentiator, having produced a preview model roughly five months after selection on around 700 government GPUs.

Outside Korea, the terrain is harsher. Qwen3.5, GLM-5.1, and DeepSeek V4 Pro Max all ship under permissive or near-permissive licenses, and as the table shows, they beat K-EXAONE 2.0 on most non-long-context, non-Korean-safety benchmarks. Their counterplay does not require any strategic creativity. It is simply cadence: release the next version sooner and cheaper. Alibaba covers the entire size spectrum so that whatever hardware you have, there is a Qwen for it. DeepSeek keeps hammering the cost-per-capability message. Against opponents shipping every few months, a once-a-year sovereign release cycle is structurally disadvantaged no matter how good any single checkpoint is.

The American labs run a third playbook worth understanding. They keep peak capability behind closed APIs while periodically releasing mid-sized open-weight models specifically to prevent developers from migrating into competing ecosystems. Layered on top, the major clouds provide optimized, one-click deployment paths for whatever they want people using. K-EXAONE 2.0 has no equivalent distribution surface today — you cannot spin it up from a console dropdown. Building that path, whether through Korean cloud providers, domestic AI accelerators, or an LG-operated API, is probably more decisive for real adoption than any benchmark on the card. LG has said a public evaluation site and a consumer-facing service are in preparation, but concrete API terms have not been announced.

And do not write off Naver. Being eliminated in the first round removed it from the program, not from the market. It retains years of accumulated work on HyperCLOVA X, a national-scale distribution surface through its own services, and — now — the freedom to ship on its own schedule rather than a ministry's. A company outside the tournament can optimize for products instead of exams. If you only track the four remaining teams, you will miss the competitor with the largest existing user base in Korea.

So What Actually Changes

If you're a developer. Do the arithmetic before you get excited. Serving the BF16 weights takes sixteen H200s per the model card, which is out of reach for individuals and for most startups, and renting that configuration by the hour is not cheap either. Two realistic paths exist. One is waiting for community quantization — there is precedent, since Q4_K_M GGUF builds appeared for the 236B first-generation model, and llama.cpp-compatible variants of that generation circulated widely. The other is waiting for a hosted endpoint, either from LG directly or from a Korean cloud provider. LG has signaled that a public evaluation site and a consumer service are coming, but nothing concrete on API pricing or availability has landed. For now, treat this as downloadable but not runnable.

If you're an enterprise buyer. This is where Apache 2.0 does the heavy lifting, and it is worth being specific about why. A house license means every deployment triggers a legal review: are there revenue thresholds, can we redistribute, who owns the fine-tuned derivative, what happens if the vendor changes terms? Apache 2.0 collapses that entire conversation, and its patent grant reduces litigation exposure as well. If you operate under data residency constraints — finance, healthcare, public sector — this belongs on your on-premises shortlist as a serious candidate rather than a curiosity. Just be honest in the total-cost-of-ownership section of your proposal. The weights are free; sixteen H200s, the power to run them, and the engineers to keep a custom vLLM fork alive are not.

If you're an investor. Do not draw a straight line from this release to LG's share price. Open-weight models generate no direct revenue by design; that is the entire point of releasing them. Three things are worth tracking instead. First, the early-August evaluation result — whether LG survives to the next round, and eventually into the final two next February. Second, real industrial and public-sector deployments, of which LG already has concrete examples: the EXAONE 4.5 vision-language model handling analysis of more than 39,000 daily public safety reports for the Ministry of the Interior and Safety, and supporting drug application reviews at the Ministry of Food and Drug Safety. Whether that reference list keeps growing is the substantive signal. Third, integration with Korean AI silicon, which would convert a model story into an industrial one. Valuing any of this off self-reported benchmark scores would be a mistake.

If you're a general user. Almost nothing changes today. K-EXAONE 2.0 is not an app, not a chatbot, and not something you can open in a browser. It is a set of weight files on Hugging Face. When LG's consumer-facing service arrives, that will change, and until then this exists for you only as numbers in headlines. The indirect effects are real but slower. If Korean services start putting this model behind their products, Korean-language response quality and cultural context handling should improve. If domestic infrastructure costs replace foreign API costs, the pricing of Korean AI services could shift too. Those are multi-quarter developments, not weekend ones.

If you're a researcher. This is the most exciting moment in the cycle for you, and also the most frustrating. Open weights for a 750B hybrid-attention MoE under Apache 2.0 means you can directly investigate why long-context retrieval nearly doubled from 52.3 to 94.4, whether the shared-expert design contributes measurably, and how upcycling redistributes capability across benchmark categories — note that MMLU-Pro slipped slightly while SWE-Bench surged, which is exactly the kind of asymmetry worth explaining. The frustration is the missing technical report and empty arXiv ID. Without training data composition and evaluation configuration, a meaningful chunk of the analysis has to wait. When that report drops, that is the real verification round.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Right now, honestly, not much. The download button works, but nobody is running a model that needs sixteen H200s on their own hardware. The realistic path to you is indirect: if Korean products start using this on the backend over the next few months, the quality and price of the Korean-language services you already use could shift quietly underneath you.

— Can I trust these benchmark numbers? Partly. The scores are published by LG under its own name on its own model card, so outright fabrication is unlikely and would be easy to catch. The problem is the evaluation setup. The model card says detailed configurations live in the technical report, that report has not been published, and the arXiv ID on the citation block is still a placeholder. Until someone independent reruns these, read them as "true under LG's methodology" rather than as settled fact.

— Will LG survive the August evaluation? Too early to call. On scale, language coverage, and safety scores, LG looks strong. But if the evaluators weight practical deployability, the sixteen-GPU serving requirement could count against it — and SK Telecom's 688B model matches it on nearly every headline spec while Upstage has explicitly positioned around running cheaply. Without published evaluation criteria, anyone claiming to know the outcome is guessing.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!