A.X-K2
A.X-K2 by SK Telecom: 688B open-weight model with Sparse Gated Attention, strong Korean and math reasoning for manufacturing, defense, and bio industries.
Model Overview
Capabilities, design details, and architectural traits
A.X-K2 - Korea's Sovereign Frontier-Scale Foundation Model
A.X-K2 is SK Telecom's 688-billion-parameter foundation model, released with open weights on Hugging Face as part of Korea's sovereign AI strategy. Built entirely domestically by a telecom operator rather than a dedicated AI lab, it succeeds the 519B A.X K1 and targets Korean industrial independence from foreign closed APIs.
The model's technical centerpiece is Sparse Gated Attention (SGA), a proprietary architecture that selectively references the most relevant information as it processes long contexts, optimizing both accuracy and operational efficiency in industrial environments. SK Telecom conducted four rounds of post-training to verify response quality, stability, and efficiency.
| Trait | Detail |
|---|---|
| Sparse Gated Attention (SGA) | Proprietary architecture that selectively references relevant information during long-context processing for accuracy and efficiency |
| 688B open-weight sovereign model | Frontier-scale parameters released openly on Hugging Face, built from scratch for Korean AI sovereignty |
| Korean-language and math specialization | Strong gains on KMMLU-Pro, CLIcK (Korean), AIME26, Apex (math); matched or beat Qwen 3.5-397B and DeepSeek-V4-Flash |
| Derivative multimodal lineup | A.X K2 VL Light-Preview (21B, vision-language), A.X K2 ALM (22B, audio), A.X K2 Raon-Speech (21B, with Krafton) |
| Industrial deployment targets | Manufacturing (KG Steel, KONEC), defense (quantized models for military units), bio (SK Biopharmaceuticals drug research) |
Industrial AI and Derivative Models
A.X-K2 is paired with a derivative lineup built around specific industry needs. The VL Light-Preview combines image and text understanding for enterprise workflows involving reports, manuals, and field imagery. The ALM recognizes and analyzes audio from customer consultations and meetings. Raon-Speech, developed with Krafton, targets voice-based player interactions in games.
The model also powers consumer and enterprise services: A. Biz for enterprise productivity tasks, and the A. (A-DoT) consumer app where it drives schedule recommendations from call conversations and content summarization.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | SK Telecom | Anthropic | Anthropic |
| Release Date | August 12, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 131K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | Free Best Input Pricing | $4 | $2 |
| Output Pricing | Free Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | text | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 21.5 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 38.8 | - | - |
| Agentic Index | 16.2 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Top AI Models
Leading alternatives by intelligence score
