This technical report presents K-EXAONE 2.0, an open-weight multilingual foundation model developed by LG AI Research as a step in our effort toward global frontier-scale foundation models. Rather than training from scratch, we upcycle K-EXAONE and expand its architecture, yielding a Mixture-of-Experts (MoE) model with 750B total parameters and approximately 37B activated per token---more than three times the capacity of its predecessor. K-EXAONE 2.0 supports context lengths of up to 256K tokens and expands multilingual coverage from six to ten languages. Its training pipeline combines continual pre-training, difficulty-focused mid-training, and post-training to strengthen reasoning, agentic coding, multilingual capability, and safety grounded in Korean sociocultural contexts. Across nine evaluation categories selected to reflect the conditions of practical use, K-EXAONE 2.0 improves over K-EXAONE and remains competitive with open-weight models, showing its largest gains in agentic coding and long-context understanding and its clearest strengths in long-context retrieval and safety. Released under the Apache 2.0 license, K-EXAONE 2.0 enables the wider AI ecosystem to evaluate, deploy, adapt, and build upon it, while marking the beginning---rather than the endpoint---of our challenge toward the global frontier.
Alex: Welcome to another episode of ResearchPod. Today, Sam, we're looking at a new technical report from LG AI Research about a model called K-EXAONE 2.0.
Sam: This paper tackles a problem that matters a lot right now: how do you build a powerful, frontier-scale AI system without the enormous budgets that only the biggest tech companies in the world can afford? The central claim is that by "upcycling" an existing, smaller model—essentially renovating it rather than demolishing it and starting over—you can reach high-level performance far more efficiently.
Alex: So the core question is: how can a smaller organization, or even a whole country, build a world-class AI without starting from scratch every single time?
Sam: Exactly. And the problem with starting from scratch is twofold. It's enormously expensive, and it throws away everything the previous model already learned. This team instead expanded their existing model's depth and capacity, keeping the foundation intact while adding significant new power on top.
Alex: It's like adding floors to a building rather than tearing it down to build a skyscraper. How does the upcycling actually work in practice?
Sam: Think of the model as a team of specialized workers—each one an "expert" in a particular kind of task. When you ask the model a question, only a small group of those experts gets called in to help. To make the model more capable, the team duplicated those experts. But here's the catch: if you just copy them exactly, the copies do the exact same work as the originals. Nothing is gained. So they added a tiny amount of random noise to each copy—a small nudge—which breaks that sameness and lets each new expert gradually learn something different.
Alex: So you're not just making more workers, you're making sure they don't all end up doing the same job.
Sam: Right. And this is where the efficiency story gets interesting. The total number of internal settings—what researchers call "parameters," which you can think of as the model's memory and knowledge capacity—grew from 236 billion to 750 billion. That sounds like a massive leap. But for any single question, the number of experts actually doing the work only grew from roughly 23 billion to 37 billion. So you have a much larger brain overall, but it doesn't burn proportionally more energy on each individual thought.
Alex: That's a meaningful distinction. But I'd imagine that adding all those extra layers and experts creates instability. How do they stop the whole thing from going haywire during training?
Sam: That's a genuine engineering challenge. As they added depth, some parts of the model started producing numbers that were far too large—runaway values that would destabilize the whole system. Their fix was a technique they call "clamped SwiGLU." Think of it like a pressure valve on a pipe. Without it, pressure builds until something bursts. With it, the system automatically releases excess pressure, keeping the numbers in a safe, predictable range throughout training.
Alex: So the valve keeps the math from spiraling. Once they have this large, stable model, what do they actually use it for?
Sam: One of their main targets is what they call "agentic coding"—and it's worth unpacking what that means. Rather than just answering a coding question in isolation, the model is designed to behave more like a working software engineer. It reads through an entire codebase, understands how the pieces fit together, identifies where something has gone wrong, and proposes a fix. To do that, it needs to hold a huge amount of information in mind at once. The paper reports a context window of up to 256,000 tokens—think of a token as roughly a word or a short phrase. That's enough to keep an entire large software project in view simultaneously.
Alex: So it's not just about raw size. It's about directing that capacity toward tasks that actually matter in the real world.
Sam: That's a fair summary. And that same logic carries into the second major part of this report, which is about safety.
Alex: Right—because a more capable model is also a more potentially risky one. How do they approach that?
Sam: The team's argument is that if you rely on safety standards built entirely in Western contexts, you'll miss risks that are specific to other cultures and languages. So they developed their own framework, which they call the Korea-Augmented Universal Taxonomy, or K-AUT. It's essentially a custom rulebook for what the model should and shouldn't do, grounded in Korean and broader Asian social contexts.
Alex: And how do they decide what goes into that rulebook without just guessing?
Sam: They brought in 46 experienced teachers, recruited in partnership with the UNESCO Asia-Pacific Centre of Education for International Understanding. These aren't engineers writing rules in the abstract—they're professionals who spend their working lives teaching global citizenship, human rights, and social norms to students. Over four weeks, these teachers tried to provoke the model into producing harmful responses, then proposed revisions to existing rules and identified 70 entirely new risk categories the original framework had missed.
Alex: That's a genuinely different approach. Instead of having a technical team define harm from the inside, you're bringing in people whose whole job is thinking about ethics in practice.
Sam: And they treat disagreement among the teachers as useful information, not a problem to be resolved quickly. When teachers diverge on whether something is harmful, that divergence itself flags a genuinely sensitive area. Multiple teachers review each response, and the research team makes the final call on which judgments become formal criteria. It's a structured way to turn human moral reasoning into something a machine can be trained on.
Alex: Does it actually change how the model behaves, though? Writing a rule is one thing.
Sam: They verify it empirically. New training data is built from those rules, the model is retrained, and then it's tested again—using human evaluators for the most sensitive cases and automated tools for speed. The paper reports a meaningful improvement in what they call the "Safe Rate" across their safety categories compared to their previous model.
Alex: So the whole project—the architecture, the training, the safety work—is really about building an AI that reflects the specific context it comes from, rather than just importing a generic global standard.
Sam: That's the core of their strategy. And they're candid about one structural limitation: you can only test for risks that someone has already thought to define. The next challenge, as they see it, is finding ways to surface risks that no one has even considered yet. That's a harder problem, and this paper doesn't claim to solve it.
Alex: A powerful system built on an existing foundation, stabilized through careful engineering, and guided by a safety process that takes cultural context seriously. Thanks for walking us through it, Sam. And thanks to everyone listening to ResearchPod.