SK Telecom A.X K2: Korea's 688-Billion-Parameter Open-Weight Model, Explained

SK Telecom's A.X K2 is one of the most significant "sovereign AI" releases of 2026: a 688-billion-parameter foundation model, released with open weights on Hugging Face, built by Korea's largest telecom operator to strengthen math and science reasoning, Korean-language knowledge, and long-context understanding. It's the successor to the 519B A.X K1 — Korea's first 500B-scale hyperscale model — and it doubles down on a distinctly national strategy: a frontier-scale model built from scratch, owned and controlled domestically, and pointed squarely at industry. This guide explains what A.X K2 actually is, what's new, what SK Telecom claims, and how to think about it against the global open-weight field.

Every figure below is labeled by source. A.X K2's benchmark gains are SK Telecom-reported (largely measured relative to A.X K1) and not independently audited; competitor figures are noted where sourced. Benchmarks and availability move — verify before building on them.

TL;DR. A.X K2 scales SK Telecom's A.X line from 519B to 688B parameters and ships open-weight on Hugging Face. SK Telecom reports average performance across 14 Korean and international benchmarks up 32.2 percentage points over A.X K1, with long-context and agent evaluations up about 83.9 points, and says it matches or beats recent Alibaba Qwen and DeepSeek releases on math and Korean tasks. The technical centerpiece is SK Telecom's own Sparse Gate Attention (SGA) architecture for efficient long-context reasoning. It's aimed at manufacturing, defense, and bio, with a longer-term roadmap toward a trillion parameters.

Key takeaways

• 688B parameters, open weights. A frontier-scale Korean model released openly on Hugging Face, up from A.X K1's 519B.

• Big reported gains over K1. +32.2 pp average across 14 benchmarks; +~83.9 pp on long-context and agent evaluations (SK Telecom-reported).

• New architecture: Sparse Gate Attention (SGA). SK Telecom's in-house design for accurate, efficient long-context processing.

• Strong on Korean and math. SK Telecom says A.X K2 matched or bested recent Qwen and DeepSeek releases on math and Korean-language benchmarks.

• Sovereign + industrial. Built from scratch for Korean AI sovereignty, targeting manufacturing, defense, and bio, with a trillion-parameter roadmap.

What A.X K2 actually is

A.X (pronounced "A dot X") is SK Telecom's proprietary large-language-model line. A.X K2 is its new flagship: a 688-billion-parameter foundation model that SK Telecom released with open weights on Hugging Face. That combination — frontier scale plus open weights, from a telecom operator rather than a dedicated AI lab — is unusual, and it's the point. SK Telecom has framed the A.X program as building a Korean "sovereign" AI: a hyperscale model developed domestically from scratch, so that Korea's strategic industries aren't dependent on foreign closed APIs. A.X K1 (519B) established that foundation in 2025 as Korea's first 500B-scale model; A.X K2 scales it up and sharpens its focus.

What's new versus A.X K1

The jump from A.X K1 to A.X K2 is both size and capability. Parameters rise from 519B to 688B, and SK Telecom concentrated the gains on three areas: mathematical and scientific reasoning, Korean-language knowledge, and long-context reasoning. The reported results are large: average performance across 14 domestic and international benchmarks improved by 32.2 percentage points over A.X K1, and — most strikingly — long-context understanding and agent-related evaluations improved by about 83.9 percentage points. Those are SK Telecom's own figures, measured against its predecessor, so read them as a vendor-reported generational leap rather than an independent ranking. For context on the baseline, A.X K1 posted 80.2 on KMMLU (the Korean counterpart to MMLU), underscoring the line's Korean-language strength.

The architecture: Sparse Gate Attention (SGA)

The headline technical contribution is SK Telecom's proprietary Sparse Gate Attention (SGA). In plain terms, SGA is designed to let the model selectively reference only the highly relevant parts of a long context instead of attending to everything equally. The goal is to get both the accuracy and the operational efficiency that industrial deployments demand — long documents, logs, and multi-step agent tasks processed without runaway compute. The outsized reported gains on long-context and agent evaluations (that +83.9-point figure) are attributed to SGA, which makes sense: those are exactly the workloads a smarter long-context attention mechanism should help most.

What SK Telecom claims on benchmarks — and the caveats

Beyond the generational gains over K1, SK Telecom says A.X K2 matched or bested recent releases from Alibaba's Qwen and DeepSeek on benchmarks covering mathematics and Korean language. That's a meaningful claim — Qwen and DeepSeek are among the strongest open/open-leaning models globally — but it comes with important caveats. The comparisons are SK Telecom-run, they focus on specific domains (math and Korean) rather than a broad, independent composite, and there's no third-party Intelligence-Index-style score for A.X K2 yet. So the fair reading is: A.X K2 is a genuinely strong open-weight model, especially in Korean and math, per its maker — but "matched or bested Qwen and DeepSeek" is a domain-specific vendor claim, not a global crown.

Open weights: how to get it

A.X K2 is open-weight and available on Hugging Face (search "SK Telecom A.X K2", repo skt/A.X-K2), just as A.X K1 was. That means teams can download, inspect, fine-tune, and self-host it — valuable for organizations with data-residency or sovereignty requirements, which is much of A.X K2's intended audience. As with any 688B model, "open" means you can self-host, not that it's cheap or trivial to serve; running a model this large takes serious hardware, so many teams will access frontier open models through hosted providers instead.

Why the industrial and sovereign framing matters

A.X K2 isn't chasing chatbot leaderboards; it's aimed at industry. SK Telecom is expanding A.X K2 use cases in manufacturing, defense, and bio — for example developing specialized AI agents with steelmaker KG Steel and with Connec, and piloting the model on a cold-rolling line and on casting/machining processes. Combined with the open-weight, built-from-scratch, domestically-controlled approach, this positions A.X K2 as sovereign industrial AI: a model a nation's strategic sectors can run on their own terms. That's a different value proposition from a general-purpose Western frontier model, and it's the lens through which A.X K2 is best judged.

The roadmap

SK Telecom has signaled that A.X K2 is a step, not the destination: its stated longer-term plan is to scale the A.X line into the trillion-parameter range. If K1 (519B) → K2 (688B) is the trajectory, a trillion-parameter A.X would be a major statement about Korea's intent to compete at the very top of open-weight scale.

How to use it alongside other models

Because the model landscape is fragmented and fast-moving, it pays to keep your integration vendor-neutral. a href="https://www.orcarouter.ai/">OrcaRouter/a> provides one OpenAI-compatible endpoint across a large model catalog — including strong open and open-leaning models like DeepSeek V4 Pro, Qwen 3.8, and Kimi K3 that A.X K2 is naturally compared against. A.X K2 itself is open-weight and self-hostable rather than a hosted API model, so the practical pattern for many teams is: self-host A.X K2 where Korean-language, sovereignty, or industrial requirements demand it, and reach the other frontier models through one vendor-neutral endpoint — so adding or switching models stays a configuration change rather than a re-integration.

Three real-world scenarios

1. Korean-language and domestic knowledge work

A.X K2's Korean strength (the A.X line's KMMLU pedigree) makes it a natural fit for Korean enterprise, government, and consumer applications where language and local knowledge matter most.

2. Long-context and agentic industrial tasks

SGA and the reported +83.9-point long-context/agent gains target exactly the document-heavy, multi-step workflows found in manufacturing, defense, and bio.

3. Sovereign, self-hosted deployment

Open weights let regulated or sovereignty-conscious organizations run a frontier-scale model entirely in-house — the core reason a telecom built it from scratch.

When not to reach for A.X K2

Don't assume it tops global general-purpose leaderboards — SK Telecom's strongest claims are domain-specific (Korean, math) and vendor-run, and there's no independent composite score yet. Don't count on cheap, turnkey access: it's a 688B open-weight model, so serving it takes real infrastructure. And if your workload is English-first general chat with no sovereignty or industrial angle, a hosted Western frontier model may be simpler. Treat A.X K2's reported numbers as provisional until third parties test it.

FAQ

What is SK Telecom A.X K2?

A 688-billion-parameter, open-weight foundation model from SK Telecom, released on Hugging Face in 2026 — the successor to the 519B A.X K1 — focused on math/science reasoning, Korean knowledge, and long-context understanding.

How is A.X K2 different from A.X K1?

It's larger (688B vs 519B) and, per SK Telecom, averages 32.2 points higher across 14 benchmarks, with long-context and agent evaluations up about 83.9 points, powered by the new Sparse Gate Attention architecture.

Is A.X K2 open source?

It's released with open weights on Hugging Face, so you can download, fine-tune, and self-host it. Check the repository for the exact license terms.

Is A.X K2 better than Qwen or DeepSeek?

SK Telecom says it matched or bested recent Qwen and DeepSeek releases specifically on math and Korean-language benchmarks. That's a domain-specific, vendor-run claim — not an independent global ranking.

What is Sparse Gate Attention (SGA)?

SK Telecom's in-house attention design that selectively focuses on the most relevant parts of a long context, aiming for both accuracy and efficiency in industrial workloads.

What is A.X K2 used for?

SK Telecom is targeting manufacturing, defense, and bio, with early work alongside partners like KG Steel and Connec, as part of a Korean sovereign-AI and industrial strategy.

Bottom line

A.X K2 is a statement of intent: a 688-billion-parameter, open-weight, built-from-scratch Korean model that SK Telecom says leaps well past A.X K1 (+32.2 points on average, +~83.9 on long-context and agents) and holds its own against Qwen and DeepSeek on math and Korean — with a trillion-parameter future in view. The numbers are vendor-reported and the model is heavy to self-host, so temper the hype until independent tests land. But as sovereign, industrial, open-weight AI, A.X K2 is one of 2026's most interesting releases. Evaluate it on your own Korean-language and long-context workloads, and keep the rest of your model access vendor-neutral through one OpenAI-compatible endpoint at OrcaRouter.