A.X K2 debuts as Korea’s largest open LLM with 688B parameters
teortaxesTex · x · 2026-07-29
A.X K2 launches as Korea’s largest open LLM with 688B total parameters
A.X K2 has been released as the largest open Korean LLM to date and another result of the Korean government’s Sovereign AI foundation model project.
The model is described as a Mixture-of-Experts system with:
- 688B total parameters and 33B active parameters
- a Think-Fusion training recipe that supports both thinking and non-thinking modes
- Gated Transformer Blocks with Sparse Gated Attention and Gated Norm for stable training and efficient inference
The post also shows benchmark comparisons versus models including A.X K1, Qwen3.5, Nemotron 3 Ultra, DeepSeek-V4 Flash, and GLM-5. In the shown table, A.X K2 leads on several tasks such as AIME26, Apex, KMMU-Pro, CLICK, LiveCodeBench v6, and τ²-Bench, while other models still win on some individual rows like KMO26 and KoBALT.
Related event: South Korea Releases A.X K2, a 688B-Parameter Open-Source Model(2 posts)→
More from Models
- Critique on Quantization Benchmarks: When Quant Beats Reference, It's Overfitting — teortaxesTex · 2026-08-24
- Gemini 3.X Infers Aug 24 as 2024 Due to Knowledge Cutoff — bclavie · 2026-08-24
- Claude's invisible watermarks cracked within hours; override code gets 20k bookmarks — deliprao · 2026-08-24
- Sonnet 4.5 exhibits intense, strange behavior in response to Opus 3 — repligate · 2026-08-24
- A comprehensive ranking of various AI models has been shared — FinanceYF5 · 2026-08-24
- User comparison finds LTX outperforms H3 in instrument generation energy — cocktailpeanut · 2026-08-24