The government's initiative to Proprietary AI Foundation Model The competition for the project has narrowed down to three teams.
According to the results of the second-stage evaluation released by the Ministry of Science and ICT on August 18, LG AI Research, SK Telecom, and Upstage...advanced to the next round, while Motif Technologies was eliminated.
There’s one particularly striking aspect of these results. According to the AAII, a global evaluation metric, Motif received the highest score among the four teams, yet it failed to make the final three in the overall evaluation.
In the end, this competition wasn't just a test to determine "which model had the highest benchmark score." From actual usability and industrial applicability to the impact on the domestic AI ecosystem You could say it was a comprehensive exhibition meant to be viewed together.
| The AI competition has narrowed from four teams to three
The "Independent AI Foundation Model" project is an initiative aimed at establishing an AI infrastructure that allows Korea to control and utilize core technologies domestically, rather than relying solely on models from overseas Big Tech companies.
This second evaluation was conducted by combining 40 points from benchmarks, 35 points from experts, and 25 points from users. Within the benchmark category, the AAII assessment by the global evaluation agency Artificial Analysis accounted for 25 points, while the evaluation by the Korea Intelligence Society Promotion Agency accounted for 15 points.
|
Assessment Areas |
Point Allocation |
Key Evaluation Points |
|
Benchmark |
40 points |
25 points from the AAII and 15 points from the NIA's internal evaluation |
|
Expert |
35 points |
Development Technologies and Achievements, Future Plans, and Impact on the Ecosystem |
|
User |
25 points |
Usability and utility as perceived by expert users and the general public |
Ten external AI experts participated in the expert evaluation, while the user evaluation involved 49 AI experts and 185 members of the general public who actually tried out the service.
The average scores for the four teams were 22.5 points for the benchmark, 28.8 points for the experts, and 17.6 points for the users. The gaps between first and fourth place were 4.0 points, 2.4 points, and 5.0 points, respectively, with the largest gap occurring in the user evaluations.
The government did not disclose the rankings for individual categories or the overall scores for each team. However, it revealed that the top-ranked teams in the three evaluation areas were all different, The fact that it was a close contest that was difficult to win based on just one strengthYou can check that.
| Why Did the AAII No. 1 Motif Get Eliminated?
Before the evaluation, the AAII score was the most talked-about figure.
|
Domestic Rankings |
Model |
AAII Score |
|
1 |
Motif 3 |
47 points |
|
2 |
Solar Open2 |
37 points |
|
3 |
A.X K2 |
35 points |
|
4 |
K-EXAONE 2.0 |
31 points |
Motif's Motif 3 scored 47 points, the highest among the four models. However, the AAII was just one factor out of a total of 100 points, and it did not determine the final ranking on its own.
The Ministry of Science and ICT explained that while Motif demonstrated outstanding performance in terms of model capabilities, it received lower ratings than other teams in terms of usability and practicality. This means that the decision to eliminate the team was not based on a single criterion, but rather on the combined results of benchmark tests, expert evaluations, and user assessments.
However, since the detailed scores for each team haven't been released, we can't say for certain exactly how many points Motif trailed in each category. What is clear is that coming in first in the benchmark doesn't necessarily mean coming in first in the overall evaluation.
In fact, the fact that Motif 3 developed its own architecture, tokenizer, optimizer, and kernel and achieved strong results at AAII stands as a testament to the technical capabilities of domestic AI startups.
| LG, SKT, and Upstage: What Are the Differences Between the Three Models?
All three teams have built massive language models, but their competitive strategies differ.
|
Developer |
Model |
Total and Active Parameters |
Key Features |
|
LG AI Research |
K-EXAONE 2.0 |
750B · approx. 37B |
Large-scale models, agentic coding, long-text understanding, safety |
|
SK Telecom |
A.X K2 |
688B · approx. 33B |
Mathematics and Korean Reasoning; Applications in Manufacturing, Defense, Biotechnology, and Other Industries |
|
Upstage |
Solar Open2 |
250B · approx. 15B |
1 Million-Token Context Window, Agent Tasks, High Inference Efficiency |
Here, "active parameters" refers to the number of parameters actually used when processing a single query. Since all three models utilize a MoE architecture that selects only a subset of experts appropriate for the input, having a large total number of parameters does not mean that all computations are performed every time.
LG AI Research highlighted its scale and comprehensive capabilities. K-EXAONE 2.0 activates a total of 750 billion parameters, with approximately 37 billion parameters per token. It supports contexts of up to 256K tokens and 10 languages, and has enhanced agentic coding, long-text understanding, and safety. Released under the Apache 2.0 license, it has opened the door to both research and commercial applications.
SK Telecom demonstrated strengths in mathematical reasoning and industrial applications. A.X K2 has a scale of 688 billion units and scored 29 out of a possible 42 points—the gold medal threshold—in the IMO 2026 problem evaluation. It also tied for first place with a 97.1% accuracy rate in the MathArena AIME 2026. Another key competitive advantage is its ability to integrate with real-world applications in manufacturing, defense, and biotechnology, as well as with actual services like A.dot.
Upstage chose a broader context and efficiency. Solar Open2 activates only about 15 billion tokens out of a total of 250 billion and supports a context window of up to 1 million tokens. The quantization model is designed to run on just two H200 chips. The service integration with the Daum and Timely portals, as well as the pilot plan utilizing the FuriosaAI NPU, also received positive reviews.
To put it simply, LG has placed emphasis on “scale and comprehensive capabilities,” SKT on “mathematics and the industrial field,” and Upstage on “long-term perspective and efficiency.”Wow.
| 1,000 B200 cards per team; changes to the scale of the competition
The three teams that have advanced to the next round will each receive approximately 1,000 NVIDIA B200 GPUs. This is an increase from the 768 GPUs per team provided in the first half of the year, and the total for the three teams combined will reach approximately 3,000.
The Ministry of Science and ICT estimated that the cost of leasing 1,000 B200 units for six months would be approximately 40 billion won per team. This means that the computing resources allocated to the three teams alone total about 120 billion won.
The expansion of GPU support also means that the remaining competition is not merely a phase of fine-tuning. The three teams are conducting more training and fine-tuning, focusing not only on model performance but also on Speed, Cost, Stability, and Service Deployment CapabilityI have to bring it up to that point.
The government plans to conduct the third evaluation as early as late 2026 through early 2027 to select the final two teams out of the three. However, it is currently reviewing whether to maintain the current competition and support system beyond 2027, so the project structure may change after the final selection.
| Now, rather than the “smartest model,” “Models that can be used'Competition'
These results show that the criteria for evaluating the domestic AI competition are changing.
While a model’s raw performance is still important, high benchmark scores don’t fully reflect the quality of the actual service. Even with the same performance, the number of GPUs required and operating costs may vary, and you’ll need to separately verify how reliably it operates with Korean documentation and in industrial settings.
There are three key points to watch moving forward. We need to look at how much the model’s inference performance can be improved, how efficiently it can operate with limited computing resources, and how quickly it can be integrated into actual businesses and public services.
It remains to be seen which combination—LG AI Research’s large-scale model strategy, SK Telecom’s industry-specific strategy, or Upstage’s efficiency-focused strategy—will secure the final two spots.
However, this evaluation—in which even the AAII’s top-ranked model could have been eliminated—made one thing clear: South Korea’s independent AI competition has now moved beyond mere scorecards to the stage of selecting models that can actually function and be widely adopted.
References
#ReaderAI #BookClub #LG_AI_Research_Institute #SK Telecom #Upstage #K_EXAONE #AX_K2 #Solar_Open2 #SovereignAI #Artificial Intelligence
