LLM Post-training (Reasoning SFT + RL)
A.X K1
SK Telecom's 519B-parameter Mixture-of-Experts LLM trained from scratch, with a single model that switches between thinking and non-thinking modes on request.
- June 2025 – November 2025
- SK Telecom
- LLM Post-training (Reasoning SFT + RL)
- LLM
- MoE
- Reasoning
- Post-training
- Reinforcement Learning
- Korean
- Open Source
My Contributions
Worked on post-training as a core contributor.
- Validated the feasibility of DAPO-based reinforcement learning on instruction-following tasks, the domain the final RL stage was built on.
- Synthesized training data for agentic coding and confirmed through SFT that it improves the model’s coding performance.
What is A.X K1?
A.X K1 is a 519B-parameter Mixture-of-Experts language model with 33B active parameters, pre-trained from scratch on about 10T tokens. It supports a 128K context and uses a 160K vocabulary sized for Korean efficiency.
The model is built to close the gap between reasoning capability and inference cost. Instead of shipping separate reasoning and chat models, A.X K1 lets the user decide per request whether the model should reason step by step or answer directly.
Post-training Recipe
1. Dual-track SFT. Two models are fine-tuned in parallel. Instruct SFT uses 2.91M samples of direct responses, and reasoning SFT uses 4.60M samples with explicit reasoning traces. About half of the reasoning data is Korean.
2. Think-Fusion. The two models are merged linearly, weighted 0.8 toward the reasoning model so that reasoning depth is preserved. The merged model is then fine-tuned on a Mode-Overlap Dataset, where each prompt has both a thinking and a non-thinking response. This step resolves the mode confusion that merging alone leaves behind.
3. On-policy RL. The fused model is trained further with a sequence-level objective and a reward that combines correctness and format validity, so that mode switching stays reliable while accuracy improves.
Sovereign AI Foundation Model Project
A.X K1 was the SK Telecom team’s model in Korea’s Sovereign AI Foundation Model project, led by the Ministry of Science and ICT. In the first-round evaluation the ministry announced on January 15, 2026, A.X K1 ranked 2nd among the five selected teams and advanced to the second phase, while two teams were eliminated. The evaluation combined benchmark tests (40 points), expert review (35 points), and user evaluation (25 points).
Results
In thinking mode, A.X K1 is competitive with DeepSeek-V3.1 (685B-A37B) and GLM-4.6 (357B-A32B), with a clear lead on math and Korean reasoning.
- Math — 89.8 on AIME25, ahead of DeepSeek-V3.1 (88.4) and GLM-4.6 (86.0). On the Korean version, AIME25-ko, the gap widens to 86.9 vs 78.1 and 75.9.
- Korean math — 89.4 on HRM8K, against 84.3 and 83.9.
- Instruction following — 64.7 on IFBench, against 41.5 and 43.4.
- Korean code — 73.1 on LiveCodeBench-ko, against 66.2 and 55.9.
- Korean knowledge — 80.2 on KMMLU, against 76.5 and 79.9.
The report also notes where the model falls behind: long-context benchmarks and complex coding, which it attributes to the smaller pre-training scale and compute budget.