A.X K2 packs 688 billion parameters and open weights
In sum – what we know:
- Open weights, national backing – A.X K2 launched on Hugging Face with weights and configs, as the centerpiece of South Korea’s government-backed independent foundation model program.
- Efficiency-focused architecture – SK Telecom added three proprietary modifications, including Sparse Gated Attention, that prune irrelevant tokens to make long-context processing cheaper.
- Unanswered compute costs – The energy requirements and environmental costs of running a 688B-parameter model remain undocumented, a potential barrier for the smaller Korean organizations it’s meant to serve.
SK Telecom has unveiled A.X K2, a 688-billion-parameter large language model that the company is positioning as South Korea’s flagship “sovereign” AI foundation model — a homegrown alternative meant to reduce the country’s reliance on foreign platforms. The model landed on Hugging Face, complete with weights and configuration details, which puts it in the same open-release camp as the likes of DeepSeek and Qwen rather than behind an API.
The launch isn’t happening in a vacuum. A.X K2 is the centerpiece of South Korea’s government-backed “independent AI foundation model” program and feeds into the broader national “AI for All” initiative, which aims to put free AI services in the hands of ordinary citizens.
Architecture and scale
At 688 billion parameters, A.X K2 sits among the larger publicly released Mixture-of-Experts models, though it’s not uniquely huge — Kimi K2 (no relation, despite the name) and other trillion-parameter MoE designs have already pushed past that mark. Like most models at this scale, only a fraction of experts activate per token, which keeps inference costs from being completely absurd.
SK Telecom’s differentiation comes in three proprietary modifications aimed at long-context efficiency. Sparse Gated Attention uses what the company calls a “lightweight indexer” in every layer to score positions, select the top candidates, and prune irrelevant tokens — essentially teaching the model to skim rather than read every word of a long document. Gated Norm and an attention output gate round out the changes, both targeting training stability and efficiency.
Performance and benchmarks
SK Telecom’s headline claim is a 32.2 percentage-point average improvement across 14 benchmarks compared with its predecessor, A.X K1 — the 519-billion-parameter model it released in the program’s first phase. The company also cites an 83.9 percentage-point jump in long-context and agent evaluations, along with substantially stronger mathematical and scientific reasoning. The comparison that matters more is the external one. SK Telecom claims A.X K2 performs comparably to Qwen3.5 and DeepSeek-V4 in reasoning and agentic tasks.
One genuinely distinctive element is the Korea-specific knowledge work. The training emphasized accurate Korean-language datasets focused on education and STEM, which is precisely the kind of thing global models handle inconsistently.
Industrial use
SK Telecom is pitching A.X K2 as a working tool rather than a research artifact, with deployments planned across manufacturing, defense, and bio/healthcare — all accessed through natural-language interfaces layered on top of existing enterprise systems.
The manufacturing use cases are the familiar ones. Process optimization, quality inspection, and predictive maintenance, with the model acting as a conversational layer over production data. Defense is the more sensitive deployment — intelligence analysis, simulation support, and multilingual translation, run under tight security with custom fine-tuning on defense-relevant data.
In bio and healthcare, the plans cover literature mining, drug discovery support, and clinical documentation summarization, though no specific deployments have been named yet. K2 is also headed into corporate tools for document drafting, meeting summarization, and internal knowledge search — the unglamorous work that tends to be where enterprise AI actually earns its keep.
Consumer applications and multimodal derivatives
For everyday users, a lightweight variant of K2 already powers features in A., SK Telecom’s AI assistant, handling call summarization and automatic schedule extraction from conversations. The bigger consumer story is the planned government service. Under the “AI for All” initiative, K2 is expected to serve as the backbone for a free, nationwide civilian AI offering aimed at personal productivity and education — details of which remain fluid.
Alongside the main model, SK Telecom announced three lightweight derivatives designed to handle combinations of text, speech, and images, with A.X K2 Light serving as the base language model underneath them. The most interesting of the bunch is A.X K2 Raon-Speech, a 21-billion-parameter speech foundation model co-developed with KRAFTON AI. It pairs a 317M-parameter speech encoder and a 96M speech codec with K2 Light to take spoken input and generate both text and synthetic speech in response. The KRAFTON involvement is a useful signal that Korea’s sovereign AI push extends beyond one telecom giant.
Open-source ecosystem
The Hugging Face release includes model weights and configuration details compatible with common inference frameworks, and it slots into the Ministry of Science and ICT’s evaluation of domestic foundation models as strategic infrastructure. The goal is a self-sustaining Korean software ecosystem, and open weights are a reasonable way to seed one.
And then there’s the practical question nobody has answered. Running a 688B-parameter model, even a sparse one, demands serious compute, and the energy requirements and environmental costs are undocumented. For the smaller Korean organizations this ecosystem is supposedly being built for, that may be the biggest barrier of all. A.X K2 is a credible technical achievement and a clear statement of national intent — but the gaps between the announcement and the accountability are exactly where the follow-up reporting needs to go.
Table of Contents
