SK Telecom open-sources Korea's largest 688B-parameter AI model on Hugging Face
The AMW Read
A.X K2 meaningfully updates the Korean foundation-model player map (§2) as the largest domestic open-weight model, and the open-source distribution strategy provides a data point on scaling-law tradeoffs at the 688B-parameter level (cross.§B).
SK Telecom open-sources Korea's largest 688B-parameter AI model on Hugging Face
SK Telecom (SKT) has open-sourced its A.X K2 model on Hugging Face under the Apache 2.0 license, making it the largest open-weight model to emerge from South Korea at 688 billion total parameters. The release includes three derivative models: A.X K2 VL Light-Preview (21B), A.X K2 ALM (22B), and A.X K2 Raon-Speech, the latter developed by consortium partner Krafton. A.X K2 employs a MoE architecture with 256 experts (8 active, 33B activated parameters), a 256K-token context window, and was pretrained on 8.2 trillion tokens. SKT claims the model matches or exceeds recent open-source releases including Qwen 3.5-397B-A17B, DeepSeek-V4 Flash, GLM-5.1, and K2-K2.6 on select benchmarks, particularly in math, Korean-language reasoning, and long-document understanding.
Why it matters: This release updates Korea's position on the global foundation-model map. SKT's A.X K2 is the flagship output of South Korea's sovereign AI program (독파모), and at 688B parameters it surpasses rival domestic offerings from LG AI Research (236B), Upstage (250B), and Motif Technologies (314B). The open-weight Apache 2.0 licensing follows an emerging pattern in which state-backed or telco-led labs release large models to seed domestic ecosystems — a variant of the open-weight distribution moat that helps hyperscalers and governments avoid vendor lock-in to US or Chinese frontier labs. SKT explicitly frames A.X K2 as a foundation for industrial AI in manufacturing, defense, and bio, with named deployments at KG Steel and SK Hynix factories.
Expert take: A.X K2 is structurally competitive with top-tier open-weight models but reveals a clear gap in compute scale. At 688B total parameters (33B active), it trails the trillion-parameter frontier set by Moonshot AI's K2.6 (2.8T total) and Alibaba's Qwen 3.8 (2.3T). The 256K context window also falls short of the 1M-token threshold now standard for frontier models. SKT's open-weight strategy is atypical for a telecom operator and mirrors the pattern set by DeepSeek and Alibaba: releasing a large MoE model to attract enterprise developers and government contracts. The real test will be whether the model's Korean-language superiority and industrial partnerships can overcome its compute disadvantage against better-capitalized Chinese and US labs. Korea's entire sovereign AI bet now rests on scaling A.X K-series to trillion-parameter territory, as SKT has signaled.
#SKTelecom #A.XK2 #FoundationModels #OpenSourceAI #MoE #SovereignAI #HuggingFace
