SKT A.X

Korea's hyperscale sovereign AI — 519B-parameter MoE LLM built from scratch for Korean-first multilingual intelligence

LLMs & Chat Freemium Has API Open Source
Researched · Published · Reviewed
RECATOOLS Score
7.2 / 10
Capability
7.5
Value for money
8.5
Ease of use
6
ASEAN readiness
7
API quality
5.5
Founded
1984
HQ
Seoul, South Korea
Users
~10M MAU (A-dot consumer app, Korea)
Launched
A.X K1 unveiled December 27, 2025
Developer
SK Group (SK Telecom Co., Ltd.)

Overview

SKT A.X is the AI platform and foundation model family from SK Telecom, South Korea's largest telecom operator. Its flagship model, A.X K1, is a 519-billion-parameter Mixture-of-Experts (MoE) language model trained from scratch on roughly 10 trillion tokens, making it Korea's first 500B-scale hyperscale LLM. Released in December 2025 under Apache 2.0, A.X K1 uses a 519B-A33B architecture (33B parameters active per token) with Multi-Latent Attention, a 131,072-token context window, and a Think-Fusion mode that lets users toggle between deep chain-of-thought reasoning and low-latency response within a single model. Smaller A.X variants — including A.X 4.0 (72B, Qwen2.5-based) and the fully in-house A.X 3.1 — are also open-sourced on Hugging Face under the skt/ organisation.\n\nThe A.X ecosystem is broader than just the model. SK Telecom's A. (A-dot) consumer AI assistant, powered by A.X, reached approximately 10 million monthly active users in Korea by late 2025, offering call-summarisation, AI note-taking, real-time call translation, and fraud filtering. On the enterprise side, A. Biz delivers manufacturing and industrial AI agent solutions, while partners including Krafton (gaming AI), 42dot (on-device AI), Rebellions (domestic NPU), Seoul National University, and KAIST extend the ecosystem. Through the Global Telco AI Alliance (GTAA) — a joint venture with Deutsche Telekom, e&, Singtel, and SoftBank — A.X models are being adapted for telco-specific multilingual LLM use cases reaching an estimated 1.3 billion subscribers across 50 countries.

Advertisement

Pricing

Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.

Enterprise
Custom
Hosted or on-premise deployment via SKT's A. Biz enterprise arm
  • Contact sales ([email protected])
  • Quantized and hosted API options
  • Korean-language optimised deployments

Use cases

Building Korean-language enterprise chatbots and RAG pipelines using the open-source A.X K1 or A.X 4.0 weights Distilling specialised smaller models (sub-70B) for domain-specific tasks by using A.X K1 as the teacher model Integrating AI agent workflows — autonomous email drafting, document generation, web search — via A. Biz enterprise APIs Developing telco-vertical AI applications (fraud detection, call summarisation, customer service) through the GTAA joint venture LLM On-device AI inference on domestic Korean NPU hardware (Rebellions) using quantized A.X variants for privacy-sensitive deployments

What you can produce with SKT A.X

  • A production-ready Korean-English bilingual chatbot fine-tuned from A.X 4.0 (72B) on your domain data
  • A distilled 7B or 13B specialist model trained using A.X K1 as the teacher, ready for on-premise deployment
  • Call transcripts and AI-generated meeting summaries via A-dot's call-summary feature for Korean-language voice calls
  • An AI fraud-detection layer for telco or financial workflows, leveraging SKT's voice-phishing classification models
  • A sovereign AI proof-of-concept deployment — model + inference + data residency — using SKT's Sovereign AI Package for government or regulated-industry clients
  • Multilingual customer service agents supporting Korean, English, Chinese, Japanese, and Spanish via A.X K1's native language coverage
  • An enterprise manufacturing AI solution for anomaly detection or quality-control using A. Biz industrial AI agent APIs
Advertisement

ASEAN Perspective

SKT A.X in Southeast Asia

SK Telecom's GTAA partnership with Singapore's Singtel — whose group covers 770 million subscribers across 21 countries including Indonesia, India, and Australia — is the clearest near-term ASEAN on-ramp for A.X technology. SKT is also building an AI data centre in Vietnam (in partnership with SK Innovation, using LNG cold-energy cooling) and has named Malaysia and Singapore as target markets for its Sovereign AI Package. Data-residency is central to SKT's pitch: the Sovereign AI Package explicitly bundles domestic AIDC infrastructure with A.X K1 so that governments and enterprises can run Korean or nationally-localised LLMs without data crossing borders — a compelling argument in ASEAN markets with growing data-protection mandates. However, as of mid-2026 no A.X production service is publicly live in any Southeast Asian country, and the GTAA telco LLM supports Bahasa as a named target language but has not yet publicly benchmarked ASEAN-language performance.

RECATOOLS Verdict

SKT A.X K1 remains a genuine feat — the first 500B-plus MoE model built from scratch by a non-hyperscaler, released under Apache 2.0 with real Korean-language benchmark wins over GPT-4o (78.32 vs 72.51 on KMMLU). 2026 brought it out of the lab: the model entered Phase 2 of Korea's government-backed Sovereign AI Foundation Model Project in January, and SK Telecom signed its first industrial deployments in June, piloting A.X K1 as a 'teacher model' inside KG Steel and KONEC manufacturing lines. It still trails DeepSeek-V3 and GLM-4 on KMMLU-Pro, and there is no self-serve public API — access runs through enterprise sales. The 2025 USIM breach exposed 23 million subscribers' authentication keys; SK Telecom is now contesting the resulting $96.9M (₩134.8bn) fine in Seoul Administrative Court, a case that will run through 2026.

Independent AI-assisted assessment by RECATOOLS.

What people say

Ten million monthly active users on the A. (A-dot) consumer app is the number SK Telecom leads with, and it holds up — Korean reviewers consistently cite call summarisation, real-time translation and fraud filtering as the features that actually get used, not just demoed. Under the hood, A.X K1's 519-billion parameters (33B active per token, Apache 2.0 licensed) posted real wins on domestic benchmarks: 78.32 on KMMLU against GPT-4o's 72.51, 83.51 on CLiCK against 80.22. That's the kind of margin that gets a model cited in academic comparisons, not just press releases.

2026 turned the model from a benchmark story into a deployment one. It cleared Phase 2 of Korea's Sovereign AI Foundation Model Project in January and, this June, SK Telecom signed its first manufacturing pilots — KG Steel and auto-parts maker KONEC are testing A.X K1 as a knowledge-transfer 'teacher model' for smaller on-site systems. Complex multi-step reasoning and coding still lag DeepSeek-V3 and GLM-4 on KMMLU-Pro, and there's still no public, self-serve API — enterprise buyers go through sales, full stop.

The one shadow that won't lift quickly: the April 2025 USIM breach, which exposed authentication keys for 23 million subscribers (not the 25 million some early coverage cited) and drew a ₩134.8 billion ($96.9M) fine from Korea's privacy regulator, the largest telecom penalty it has ever issued. SK Telecom is fighting the fine in Seoul Administrative Court as of early 2026, arguing its $812M cleanup and compensation spend should count against the number. Regulated-industry buyers evaluating A.X should weigh that history alongside the model's real technical strengths.

Summary of public user & expert reviews, compiled by RECATOOLS.

Notable facts

  • A.X K1's 519B total parameters activate only 33B at a time per token — making it as computationally efficient as a ~33B dense model while retaining the knowledge capacity of a 519B one.
  • SK Telecom's A-dot AI assistant blocked approximately 1.1 billion voice-phishing and spam call attempts in 2025 alone — more than 3 million fraudulent calls stopped per day.
  • The GTAA telco alliance (SKT + Deutsche Telekom + e& + Singtel + SoftBank) collectively reaches an estimated 1.3 billion mobile subscribers across 50 countries — larger than the population of any single nation except China or India.
  • Despite being a telecom company, SKT trained A.X 3.1 entirely in-house — model architecture, data curation, and training — on its own TITAN supercomputing cluster, without relying on any external base model.

Frequently asked questions

Is A.X K1 truly open source and free to use commercially?
Yes. A.X K1 is released under Apache 2.0 on Hugging Face (huggingface.co/skt/A.X-K1), which permits commercial use, modification, and redistribution. The weights are freely downloadable. For quantized or hosted inference versions, SKT asks you to contact [email protected], and enterprise B2B access is handled through A. Biz sales rather than a self-serve API portal.
How does A.X K1 compare to DeepSeek or GPT on benchmarks?
A.X K1 is competitive but not at the very top on general benchmarks. On Korean language tasks A.X 4.0 outperforms GPT-4o (KMMLU 78.32 vs 72.51). On broader reasoning, A.X K1 scored 60.3 on KMMLU-Pro versus DeepSeek-V3's 67.9, reflecting a smaller overall training compute budget. Its strongest claim is Korean-language cultural understanding and instruction following.
What happened with SK Telecom's 2025 data breach — does it affect A.X?
In April 2025, a BPFDoor malware attack on SKT's telecom infrastructure exposed USIM authentication data for approximately 25 million subscribers. The breach was in SKT's telco core network, not the A.X LLM platform itself. SKT was fined $96.9M and offered free SIM replacements. Organisations evaluating SKT as an AI cloud or data-processing partner should factor in this incident and the subsequent regulatory findings when assessing vendor trust.

About this listing

Researched on
Published on
Last reviewed

This entry was compiled from publicly available data including SKT A.X's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with SKT A.X unless explicitly stated.

Data accuracy

Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.

For the latest details, please refer to SKT A.X directly →

Spotted something out of date? Suggest an update →

Advertisement