A.X-K2

A.X-K2 is a large-scale language model developed by SK Telecom as a central component of South Korea's "Sovereign AI" initiative. Released as the successor to the A.X-K1, the model features a Sparse Mixture-of-Experts (MoE) architecture with 688 billion total parameters, of which 33 billion are active during any single forward pass. It was trained from scratch on approximately 8.2 trillion tokens of high-quality data, with a specific focus on the Korean language, mathematics, and long-context reasoning.

The model's architecture is characterized by its Sparse Gated Attention (SGA) mechanism, an in-house design that enhances efficiency for long-sequence processing by selectively indexing and referencing the most relevant past tokens. A.X-K2 consists of 61 layers (including 60 MoE layers) and utilizes 256 routed experts alongside one shared expert. This design allows the model to maintain a 262,144-token (256K) context window with high retrieval accuracy and operational efficiency in industrial environments such as manufacturing, defense, and biotechnology.

A unique feature of A.X-K2 is its Think-Fusion training recipe, which enables a single checkpoint to support dual operating modes: a Think mode for complex reasoning and a high-latency "thinking" process, and a Non-Think mode for concise, low-latency responses. This hybrid control allows users to adjust the model's reasoning depth on a per-request basis. The model was trained natively in FP8 precision (MXFP8) for both forward and backward passes, which is reflected in the high-precision weights provided in the open-source release.

Under official benchmarks, the model demonstrated strong performance in mathematical and scientific reasoning, reaching 97.1 on AIME26 and 80.5 on KMMLU-Pro. A.X-K2 is distributed under the Apache 2.0 license, making it accessible for commercial and research applications within the global open-weight ecosystem.