South Korea's AI Foundation Project Nears Second Evaluation Amidst Intense Competition

South Korea's 'Reader AI Foundation' (Dokpamo) project is on the cusp of its second evaluation, a critical step in its mission to reduce reliance on foreign AI models and foster a robust domestic AI ecosystem. The initiative, launched in August 2025, aims to establish Korea's own AI sovereignty.

“

The Dokpamo project started with five consortia, but two were eliminated in the first phase, and one new consortium has joined, leaving four contenders for the second evaluation.

The first evaluation phase rigorously assessed model originality and language model composition, leading to the elimination of two consortia: NC AI and Naver Cloud. For the second phase, Motif Technologies has been added, bringing the total number of competing consortia to four. This phase will scrutinize not only the performance of large language models but also the practical applications of agent AI in industrial settings and the capabilities of multimodal technology, which processes diverse data types like text, images, and video simultaneously.

Upstage's Solar of Open 2 model boasts a significantly expanded parameter count of 250 billion, more than double its predecessor's 102 billion. While it uses 15 billion parameters for actual computation, it can process one million tokens at once, enabling it to handle lengthy and complex AI agent tasks without losing context. The model reuses approximately 2.3% of its core weights from the previous version, reducing training time and enhancing Korean language token efficiency. It has also been trained on agent scenarios for conversational tool use, coding, and office tasks, directly addressing the second evaluation's focus on agent performance.

“

Solar of Open 2 is configured with a total of 250 billion parameters, more than double the 102 billion parameters of its predecessor.

Performance metrics show substantial improvements across various domains, including knowledge and scientific reasoning, mathematics, coding, instruction following, and overall Korean language capabilities. The GPQA-Diamond test for doctoral-level scientific reasoning saw accuracy jump from 66.2% to 86.3%. Live Bench v6 performance improved from 56.5% to 92.4%, and the AIME2026 math test accuracy rose from 87.7% to 95.7%. Overall performance now rivals global frontier models like Deepseek-V4 Flash 284-A13B.

“

The GPQA-Diamond test for doctoral-level scientific reasoning saw accuracy jump from 66.2% to 86.3%.

SK Telecom has unveiled its A.X K2, an upgraded model with 688 billion parameters, an increase from the 519 billion parameters of its predecessor, A.X K1. This iteration includes specialized models for image analysis (A.X K2 VE), acoustics (A.X K2 ALM), and speech (A.X K2 Raon-Speech) to enhance its multimodal capabilities. The expert model's active parameters remain at 33 billion, but the structure has been enhanced from 192 experts to 256. Context length has doubled to 260,000 tokens. A.X K2 incorporates sparse-gated attention, combining sparse attention and gated attention, to extend context reasoning length while maintaining speed and stability. Performance benchmarks show a GPQA-Diamond score of 85.6% and an AIME 2026 score of 97.1%, indicating strong intelligence capabilities.

“

A.X K2 is an upgraded model with 688 billion parameters, an increase from the 519 billion parameters of its predecessor, A.X K1.