
SK Telecom A.X 4.0
Korea's top open Korean LLM: 72B model beating GPT-4o on Korean benchmarks.
About
Released July 3, 2025, A.X 4.0 is SK Telecom's open-weights Korean-optimized large language model, built by applying large-scale continual pre-training on Korean data to the Qwen2.5 architecture. It ships in two sizes: a 72B standard model and a 7B lightweight version. On KMMLU, the Korean adaptation of MMLU, it scored 78.3 versus GPT-4o's 72.5; on CLIcK, which measures Korean cultural and linguistic understanding, it scored 85.7 versus GPT-4o's 80.2. A custom Korean tokenizer gives it roughly 33% better token efficiency than GPT-4o on Korean text, and it supports 131K context (16K on the Light version). SKT uses it in production for its A. call summarization service.
Limitations: Korean-first by design — English and other-language performance is secondary and less benchmarked. The 4.0 series trades the full in-house sovereignty of the A.X 3 line (built from scratch) for performance via Qwen2.5 continual pre-training. On-premise deployment is the intended enterprise path, so casual users mostly meet it through SKT services rather than a polished chat product.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
KMMLU 78.3 (GPT-4o: 72.5); CLIcK 85.7 (GPT-4o: 80.2) — SKT-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 53; Trending score 68/100 from researcher assessment.