# A.X K1 Technical Report

Source: https://www.youtube.com/watch?v=KblrUI7X7Qo
Recap page: https://rapidrecap.app/video/KblrUI7X7Qo
Generated: 2026-02-06T21:37:14.45+00:00

---
## Quick Overview

SK Telecom's AXK1 technical report, released February 3rd, 2026, details a sovereign AI project utilizing a 519 billion parameter model that performs better than proprietary models on Korean language tasks by specifically training its custom 160k token vocabulary tokenizer and using expert merging (80% Math brain, 20% Chatbot) to handle complex reasoning, resulting in a stable signal for inference control.

**Key Points:**
- SK Telecom released the AXK1 technical report on February 3rd, 2026, detailing a sovereign AI project.
- The model boasts 519 billion parameters and performs better than proprietary models on Korean language tasks.
- Key innovations include a custom tokenizer trained on 160,000 Korean tokens and an expert merging strategy (80% Math brain, 20% Chatbot).
- The merging technique successfully avoided 'mode confusion' and produced a stable signal for inference control, outperforming the baseline model on MMLU.
- The model demonstrated superior performance in Korean tasks, achieving similar results to large models with far fewer tokens, proving the efficiency of their specialized tokenization.
- The researchers explicitly taught the model that one question can have two valid styles of response based on the desired mode (e.g., reasoning vs. instruction following).
- The overall cost of training was estimated to be comparable to building models in other countries, suggesting replicable efficiency.

![Screenshot at 00:10: The introduction of the AXK1 model name and reference to the technical report released on February 3rd, 2026.](https://ss.rapidrecap.app/screens/KblrUI7X7Qo/00-00-10.jpg)

**Context:** The video discusses the technical details of SK Telecom's proprietary large language model, AXK1, which is positioned as a national AI project for South Korea. The discussion centers on how the model achieves high performance, particularly in handling Korean language and culture, despite not relying solely on the massive parameter counts seen in global competitors. A key focus is on the architectural choices, such as expert merging and custom tokenization, that contribute to its efficiency and superior localized performance.

## Detailed Analysis

SK Telecom released its AXK1 technical report on February 3rd, 2026, detailing their sovereign AI project. This model is massive, containing 519 billion parameters, yet it is significantly more efficient than proprietary models when handling Korean language tasks. The researchers achieved this by focusing on innovations beyond raw scale. They built a custom tokenizer trained on 160,000 Korean tokens, which compresses information more efficiently than standard tokenizers optimized for English. Furthermore, they employed an expert merging technique (80% Math brain, 20% Chatbot) to combine different specialized models. This merging strategy, which avoids 'mode confusion' that plagues other combined models, allowed the model to switch between high-cost reasoning and low-cost instruction following modes dynamically. When tested on complex math problems, the merged model scored 89.8 on MMLU, beating the baseline model, proving the validity of their approach. The report emphasizes that future AI efficiency lies not just in size but in smarter architectural choices like controlled inference and culturally aware tokenization.

### AXK1 Model Overview

- Released February 3rd, 2026
- 519 billion parameters
- Sovereign AI project for South Korea

### Key Innovation - Custom Tokenizer

- Trained on 160,000 Korean tokens
- Compresses information more efficiently for Korean language/culture

### Key Innovation - Expert Merging

- 80% Math brain, 20% Chatbot model merged
- Avoids 'mode confusion' by allowing explicit control over reasoning vs. instruction following

### Performance Metrics

- Scored 89.8 on MMLU (Math benchmark) in thinking mode
- Outperformed baseline models
- Excellent performance on Korean sentence generation with fewer tokens

### Training & Cost

- Trained on 1,536 Nvidia H200 GPUs over four months
- Budget comparable to building similar models in other countries

![Screenshot at 00:00: Introductory graphic featuring two podcasters and the call to action 'Become a Member Today!' overlayed on an oscilloscope-like wave pattern.](https://ss.rapidrecap.app/screens/KblrUI7X7Qo/00-00-00.jpg)
![Screenshot at 00:23: Speaker explicitly stating the model is 'explicitly framed as a sovereign AI project for South Korea'.](https://ss.rapidrecap.app/screens/KblrUI7X7Qo/00-00-23.jpg)
![Screenshot at 00:48: Speaker mentioning the model's innovative feature, 'Think Fusion', which allows controlling inference style.](https://ss.rapidrecap.app/screens/KblrUI7X7Qo/00-00-48.jpg)
![Screenshot at 01:30: Speaker detailing the two options users face: using a cheap chatbot or a heavy-duty reasoning model.](https://ss.rapidrecap.app/screens/KblrUI7X7Qo/00-01-30.jpg)
![Screenshot at 02:21: Speaker noting that the 519 billion parameter count is misleading, as only 33 billion active parameters are used per token generation.](https://ss.rapidrecap.app/screens/KblrUI7X7Qo/00-02-21.jpg)