On July 29, 2026, SK Telecom released A.X K2, a 688-billion-parameter open-weight AI model, directly onto Hugging Face. The release is part of South Korea’s government-backed Sovereign AI Foundation Model project and targets three sectors that rarely appear in frontier AI announcements: steelmaking, national defense, and drug discovery.
For enterprise AI leaders watching the global model race, A.X K2 carries a message that goes beyond benchmark scores. A major telecommunications operator, backed by a national government and a fraction of the compute that Western labs use, has released a frontier-scale model under Apache 2.0 and immediately begun deploying it in factory cold-rolling lines and military units. Sovereign AI is no longer a theoretical risk-management concept. It is operational.
What A.X K2 Is
A.X K2 is the successor to SK Telecom’s A.X K1 (519 billion parameters) and was built as part of Korea’s Independent AI Foundation Model project, known domestically as “Dokpamo.” The model uses a Mixture-of-Experts (MoE) architecture: 688 billion total parameters, but only 33 billion are active per inference call, activated by selecting 8 of 256 routed experts per token.
The architecture makes large-scale deployment more practical than the raw parameter count suggests. SK Telecom reports that training A.X K2 on FP8 precision from the outset, without a separate quantization step, cut storage and inference costs by roughly half compared to standard FP32 or FP16 training.
The headline architectural innovation is SK Telecom’s proprietary Sparse Gate Attention (SGA) mechanism, which selectively processes only the most relevant parts of long documents instead of attending over the full sequence at every layer. SGA is the primary driver behind the model’s long-context gains, which we will cover below. The model also applies Gated Norm, a training stability technique that reduces weight instability during pre-training.
Key technical specifications:
| Attribute | A.X K2 |
|---|---|
| Total parameters | 688 billion |
| Active parameters per token | 33 billion |
| Expert configuration | 256 routed + 1 shared |
| Experts activated per token | 8 |
| Context window | 256K tokens (128K native, extended via YaRN) |
| Training compute | 512 NVIDIA B200 GPUs, ~70 days |
| Training data | 8.5 trillion tokens |
| Training precision | Native FP8 (MXFP8 E4M3) |
| License | Apache 2.0 |
| Architecture | Decoder-only Transformer with MoE |
Sources: SK Telecom Newsroom, Hugging Face model card, Edaily benchmark report
Benchmark Performance
SK Telecom reports average performance across 14 domestic and international benchmarks improved by 32.2 percentage points over A.X K1. Long-context and agent-related evaluations improved by approximately 83.9 percentage points, a gain the company attributes directly to SGA.
On domain-specific evaluations, A.X K2 outperforms recently released models from Alibaba (Qwen 3.5-397B) and DeepSeek (V4-Flash) on mathematics and Korean language:
| Benchmark | A.X K2 Score | What It Tests |
|---|---|---|
| AIME26 | 97.1 | Mathematical reasoning |
| KMMLU-Pro | 80.5 | Korean-language knowledge |
| CLIcK | 91.6 | Korean cultural and language understanding |
| tau2-Bench (Telecom) | 98.0 | Domain-specific telecom QA |
| IMO 2025 | 35 of 42 | Competition mathematics |
| KMO 2026 Round 2 | 8 of 8 | Korean Mathematical Olympiad |
These are vendor-reported numbers measured against the prior model and selected peer models. Independent third-party evaluation across the full Open LLM Leaderboard or LiveBench suite has not yet been published. Enterprise teams planning production deployments should run their own evals on workload-representative data before treating these scores as definitive.
Vertical Deployment Strategy
What separates A.X K2 from most open-weight model releases is not the architecture or benchmark table. It is where SK Telecom is actually deploying it.
Manufacturing. SK Telecom is building specialized AI agents using A.X K2 in partnership with KG Steel, with deployment planned for the Dangjin cold-rolling line in the second half of 2026. A parallel trial is underway with auto-parts manufacturer KONEC across casting and machining processes. Both programs aim to embed operational domain knowledge directly into AI agents that understand the physical processes in the factory, not generic assistants querying unstructured text.
Defense. Under the Sovereign AI Foundation Model project, SK Telecom developed three quantized models derived from A.X K1 in collaboration with South Korea’s Ministry of National Defense. These will be deployed across the Army, Navy, Air Force, and Marine Corps. The quantization step was specifically designed to reduce memory usage and increase processing speed within the hardware constraints common in military operational environments.
Biotech. In collaboration with SK Bioscience, A.X K2 is being applied to targeted cancer therapy research. SK Telecom reports that AI-assisted initial drug development research was compressed from one to two years down to approximately five months.
Enterprise productivity. The A.Biz enterprise service uses A.X K2 to generate newsletters, gather and analyze business information, and produce reports. A lightweight model variant is deployed in SK Hynix’s internal tools for document drafting and information management.
What Sovereign AI Means for Enterprise Strategy
The SK Telecom release is one data point in a pattern Enera has tracked across 2026: sovereign AI is moving from policy discussion to production reality. Upstage Solar Open 2 from another Korean lab arrived earlier this year as a self-hostable agentic model. Moonshot AI’s Kimi K3 arrived as the largest open-weight model at 2.8 trillion parameters. A.X K2 sits in a different position from either: it is built by a listed telecommunications company with a government mandate, trained on Korean-language data at scale, and immediately embedded in industrial operations.
For enterprise technology leaders, the practical implications are three.
Vendor diversification is now achievable at frontier scale. For most of the past two years, “frontier model” meant closed API: GPT-5, Claude Opus, Gemini Ultra. A.X K2 demonstrates that open-weight models are catching up to closed incumbents on domain-relevant benchmarks, not just general leaderboards. The gap between self-hostable and frontier-only is narrowing faster than most enterprise roadmaps anticipated.
Data residency and sovereignty have a real model tier now. For regulated enterprises in financial services, defense contracting, healthcare, and critical infrastructure, the ability to run a 688B model entirely on-premise, on domestic infrastructure, under Apache 2.0, changes the compliance calculus. The model can be fine-tuned, audited, and deployed without data leaving the organization’s control perimeter.
Vertical specialist agents are replacing horizontal assistants. The A.X K2 deployment pattern, specialized agents for cold-rolling lines and military communications, not generic chatbots, reflects the maturation phase of enterprise AI. General-purpose assistants provided the proof-of-concept. What enterprise leaders are buying now is domain-embedded reasoning that understands their specific operational context. The 83.9 pp improvement in agent-related evaluations suggests A.X K2 was engineered specifically for agentic, multi-step workloads, not just text generation.
The Enterprise Calculus on Open-Weight Frontier Models
A.X K2 does not solve the operational complexity of running a 688B model. SK Telecom’s own documentation notes the recommendation to use vLLM, SGLang, or TokenSpeed inference engines, and a standard SGLang deployment requires either 8 NVIDIA B300/MI350X-class GPUs, 16 B200/H200 GPUs, or 32 H100 80GB GPUs, depending on configuration. For most enterprises, the self-hosted path requires either a significant on-premise GPU cluster or a hosted provider that supports A.X K2.
The Apache 2.0 license does simplify the commercial equation considerably. There are no per-token royalties, no usage-based licensing fees, and no restrictions on fine-tuning for internal deployment. For enterprises with the compute, the cost structure of a self-hosted open-weight model at this scale differs fundamentally from closed API economics at high volume.
SK Telecom plans to scale the next A.X model to the trillion-parameter range, continuing its position as the lead competitor in South Korea’s four-team Sovereign AI project alongside LG AI Research, Upstage, and Motif Technologies.
If your enterprise is evaluating open-weight frontier models for sovereign deployment, regulated industries, or vertical AI agent programs, A.X K2 deserves a place on the evaluation list. Book a call with Enera to discuss how sovereign AI model strategy fits your enterprise AI architecture.
Sources: SK Telecom Newsroom (July 29, 2026); Hugging Face skt/A.X-K2; LightReading (July 29, 2026); Edaily (July 29, 2026); The Asia Business Daily (July 29, 2026); DigitalToday (July 29, 2026)