K-EXAONE 2.0 Ships 750B Open Weights Under Apache 2.0

LG AI Research has published K-EXAONE 2.0 on Hugging Face, a 750-billion-parameter foundation model released under Apache 2.0. It is the largest AI foundation model developed in South Korea to date, built under the country's Sovereign AI Foundation Model Project, and the licensing decision is arguably as consequential as the parameter count.

• 750 billion total parameters with roughly 37 billion active per token, using a hybrid-attention mixture-of-experts design
• 256 expert modules, with eight selected dynamically for each generated token
• Released under Apache 2.0, replacing LG AI Research's previous proprietary license and permitting commercial use
• Weights went live on Hugging Face on July 31, 2026

What Does 750B Total and 37B Active Actually Mean?

This is a sparse mixture-of-experts model, so the two numbers describe different things. The 750 billion figure is the total weight count you must store; the 37 billion figure is roughly how much of the network participates in producing any single token. A router picks eight of the 256 expert modules per token, and the rest sit idle.

The practical consequence is that inference cost tracks the active parameter count far more closely than the total. A 37-billion-parameter forward pass is a very different proposition from a dense 750-billion-parameter one, though you still need the memory to hold all the experts. We walked through this trade-off in detail in our explainer on sparse mixture of experts in 2026 models, and the same architecture underpins recent releases like Qwen3.8-Max.

The hybrid-attention element is the other design choice worth noting. Mixing attention variants within a single architecture has become a standard way to hold long-context cost down without giving up quality on the shorter spans where most tokens actually live.

Why the Apache 2.0 License Matters Most

LG AI Research previously shipped EXAONE models under a proprietary license that restricted commercial deployment. Moving to Apache 2.0 removes that friction entirely: download, modify, deploy commercially, no negotiation. For a national sovereign AI program, that is a deliberate statement — the model is infrastructure, not a product.

It also places K-EXAONE 2.0 directly alongside the open-weight releases that have defined this year, from Kimi K3's 2.8 trillion parameters onward. The competitive dynamic in open weights is no longer about who releases anything at all; it is about who releases under terms permissive enough that a company can build a business on top without a lawyer in the room.

Who Can Realistically Run It?

Honestly: not hobbyists, at least not at full precision. 750 billion parameters of stored weights is a datacenter-class memory requirement even when only 37 billion are active per token. The near-term audience is cloud providers, national research programs, and enterprises with existing GPU fleets.

The longer-term audience is everyone else, by way of quantization and distillation. Apache 2.0 is precisely the license that makes those derivative works legal to ship, and the open-weight community has consistently turned frontier releases into something that runs on far less hardware within weeks. That pipeline — frontier release, community compression, local deployment — is the pattern to watch, and we follow it in our AI coverage.

Sources: Hugging Face — July 31, 2026; The Korea Times — July 31, 2026; Korea JoongAng Daily — August 2026.