Four elite teams taking part in the government's "sovereign AI foundation model" project are heading into a second-round review with different scale-up strategies. Two large companies have secured parameter sizes similar to major global open-weight models. Two startups focus on improving compute efficiency and task performance relative to their size.

Industry sources said on Monday that Motif Technologies submitted a performance report for the project's second-round review the previous day. Along with LG AI Research, SK Telecom and Upstage, which had submitted earlier, all four teams have completed preparations for the second-round review.

All four teams adopted a mixture-of-experts (MoE) structure. The approach places multiple compute modules, known as experts, inside a model and activates only the required ones when a query comes in. This can increase overall model capacity while reducing the amount of computation needed to generate an answer.

MoE models should be assessed by looking at both total parameters, which represent overall size, and active parameters, which are actually deployed to process a query.

The two large companies released medium-to-large models with more than 600 billion total parameters. A global model in a similar size class is DeepSeek R1, with 671 billion total parameters and 37 billion active parameters. LG AI Research's K-Exaone 2.0 activates about 37 billion out of 750 billion total parameters per token, while SKT's A.X K2 activates 33 billion out of 688 billion per token.

The two startups chose models of around 300 billion parameters. Upstage's Solar Open 2 activates 15 billion out of 250 billion total parameters, and Motif 3 Preview Beta activates about 13 billion out of 314 billion. That is a similar size class to DeepSeek V4 Flash Preview, released in April, with 284 billion total parameters and 13 billion active parameters.

LG expands model 'by adding floors' with three-times larger model using knowledge from prior version

LG AI Research scaled up its model size most aggressively. K-Exaone 2.0's total parameters increased more than threefold from the first model, which had 236 billion total parameters.

LG AI Research increased the number of compute layers that process information in stages to 78 from 48 and expanded expert modules placed in each MoE layer to 256 from 128. Active parameters increased to 37 billion from 23 billion.

It chose an "upcycling" approach that uses weights learned by the existing model as the starting point for the new model. The concept is similar to adding floors and space while keeping a building's frame. LG AI Research explained that it can preserve knowledge acquired in the prior model while increasing compute depth and overall capacity.

In LG AI Research's own assessment, performance also rose overall. K-Exaone 2.0 scored an average of 70.1 across 24 metrics, more than 10 percent higher than the first model's 63.3. Average performance across three key metrics for coding and agentic coding rose by about 30 percent.

LG AI Research plans to use the model as a foothold to challenge the development of a trillion-parameter model in the next stage.

Lim Woo-hyung (임우형), co-head of research at LG AI Research, said it was meaningful that domestic researchers independently completed the entire process, from designing a 750 billion-class model to data training, distributed training and building the inference environment. He said they had secured the capability to compete in the same size class as global frontier models.

SKT builds 688 billion-parameter model while keeping active compute unchanged

SKT focused on reducing serving burdens such as compute and memory while increasing the size of its second model, A.X K2, by about 30 percent from the first. It expanded total parameters to 688 billion from 519 billion by increasing the number of experts, but kept active parameters at 33 billion per token.

It also chose quality over quantity in training data. The previous model trained on 10 trillion tokens, but A.X K2 used about 8.5 trillion. Instead, it concentrated higher-quality and more difficult data in the later stages of training.

It also reduced computation required to read long documents. SKT's "Sparse Gate Attention (SGA)" references only information with high relevance to the current query. It does not need to revisit the entire document with equal weighting, reducing processing burdens as inputs grow longer.

According to A.X K2's technical report, long-context comprehension scores remained at a similar level after applying SGA. Speed for processing long inputs increased, while the time to start an answer and the time to generate subsequent tokens fell compared with the earlier model.

SKT plans to scale follow-on models to the trillion-parameter level while also pursuing quantisation and optimisation for domestically made neural processing units (NPUs). Serving memory used by a quantised A.X K2 is about one-third that of A.X K1.

An SKT official said open-source models are increasingly exceeding 1 trillion parameters, making large-model development necessary for global competition. The official said serving-efficiency measures such as quantisation can reduce the burden that comes with larger model sizes.

Upstage emphasises 1 million-token context and real-world agent use

Upstage said Solar Open 2 supports a context window of up to 1 million tokens while lowering the cost of handling long texts. To do so, it mixed linear attention, which compresses prior context for processing, with existing softmax attention, which tightly calculates relationships between words. Upstage said it cut memory and computation required for long-context handling to about one-quarter compared with applying only the existing method across all layers.

It also highlighted its tokenizer, which breaks Korean sentences into smaller units that AI can process. In its in-house Korean office-work evaluation input, it packed 4.41 bytes of information per token, about 24 percent more Korean information than the most efficient model among the global models it compared. It said the same document can be handled with fewer tokens, cutting costs in long-running work where conversations and work logs accumulate.

Upstage also focused on strengthening AI agent capabilities to perform real work. It integrated training results from 12 specialised models trained by task, including tool use, coding and document writing. The aim was to enable completion of tasks that require multiple rounds of inference and tool calls.

In Upstage's in-house environment, Solar Open 2 averaged 85.4 across nine Korean benchmarks, compared with 84.9 for DeepSeek V4 Flash Preview. In English evaluations, the two models alternated in leading or lagging depending on categories such as knowledge and reasoning, coding and agents. Upstage said it strengthened Korean and real-world task execution capabilities at a similar parameter scale.

Upstage Chief Executive Kim Sung-hoon (김성훈) said Solar Open 2 was built with an emphasis on agent usability that completes work on its own in real workplaces, beyond simple benchmark scores.

Motif ties with Pro at V4 Flash class size, betting on in-house design

Motif highlighted model design efficiency as a differentiator. Motif 3 Beta, released ahead of its second-round review model, received a score of 44 on AAII, a composite index that global AI evaluator Artificial Analysis calculates by summing results across nine evaluations. That matched DeepSeek V4 Pro (Max), which has 1.6 trillion total parameters, about five times larger than Motif 3.

Motif 3 Beta activates about 13 billion parameters out of 314 billion total. Compared with DeepSeek V4 Pro's 49 billion active parameters, the scale used to generate answers is about one-quarter. It posted the same composite score with a smaller model and less active computation.

Motif emphasised it achieved the result with a from-scratch model trained from the beginning with an architecture designed in-house, rather than adopting a structure from overseas open-source models. The company said a research team of about 30 used about 700 GPUs to develop the preview model in five months.

The result covers an intermediate-stage model, not the final model submitted for the second-round review. Performance of the final submitted model has not yet been disclosed.

Motif Technologies Chief Executive Lim Jung-hwan (임정환) said he expects performance to improve over the beta model and described the result as an intermediate step toward global leading models. He added that the company will advance model performance through ongoing research and development and prove the competitiveness of Korean AI technology in the global market.

The government plans to select three of the four teams to advance to the next stage by combining benchmarks, expert evaluations and a user evaluation involving a 200-member public evaluation group. The public evaluation group assessment runs from Aug. 8 to 11.