2
A.X K2 packs 688 billion parameters and open weights
In sum – what we all know:
- Open weights, nationwide backing – A.X K2 launched on Hugging Face with weights and configs, because the centerpiece of South Korea’s government-backed impartial basis mannequin program.
- Effectivity-focused structure – SK Telecom added three proprietary modifications, together with Sparse Gated Consideration, that prune irrelevant tokens to make long-context processing cheaper.
- Unanswered compute prices – The power necessities and environmental prices of operating a 688B-parameter mannequin stay undocumented, a possible barrier for the smaller Korean organizations it’s meant to serve.
SK Telecom has unveiled A.X K2, a 688-billion-parameter massive language mannequin that the corporate is positioning as South Korea’s flagship “sovereign” AI basis mannequin — a homegrown different meant to scale back the nation’s reliance on overseas platforms. The mannequin landed on Hugging Face, full with weights and configuration particulars, which places it in the identical open-release camp because the likes of DeepSeek and Qwen fairly than behind an API.
The launch isn’t occurring in a vacuum. A.X K2 is the centerpiece of South Korea’s government-backed “impartial AI basis mannequin” program and feeds into the broader nationwide “AI for All” initiative, which goals to place free AI companies within the arms of extraordinary residents.
Structure and scale
At 688 billion parameters, A.X K2 sits among the many bigger publicly launched Combination-of-Consultants fashions, although it’s not uniquely enormous — Kimi K2 (no relation, regardless of the title) and different trillion-parameter MoE designs have already pushed previous that mark. Like most fashions at this scale, solely a fraction of consultants activate per token, which retains inference prices from being utterly absurd.
SK Telecom’s differentiation is available in three proprietary modifications geared toward long-context effectivity. Sparse Gated Consideration makes use of what the corporate calls a “light-weight indexer” in each layer to attain positions, choose the highest candidates, and prune irrelevant tokens — primarily instructing the mannequin to skim fairly than learn each phrase of an extended doc. Gated Norm and an consideration output gate spherical out the adjustments, each focusing on coaching stability and effectivity.
Efficiency and benchmarks
SK Telecom’s headline declare is a 32.2 percentage-point common enchancment throughout 14 benchmarks in contrast with its predecessor, A.X K1 — the 519-billion-parameter mannequin it launched in this system’s first part. The corporate additionally cites an 83.9 percentage-point bounce in long-context and agent evaluations, together with considerably stronger mathematical and scientific reasoning. The comparability that issues extra is the exterior one. SK Telecom claims A.X K2 performs comparably to Qwen3.5 and DeepSeek-V4 in reasoning and agentic duties.
One genuinely distinctive ingredient is the Korea-specific data work. The coaching emphasised correct Korean-language datasets centered on training and STEM, which is exactly the type of factor world fashions deal with inconsistently.
Industrial use
SK Telecom is pitching A.X K2 as a working instrument fairly than a analysis artifact, with deployments deliberate throughout manufacturing, protection, and bio/healthcare — all accessed by natural-language interfaces layered on prime of present enterprise methods.
The manufacturing use circumstances are the acquainted ones. Course of optimization, high quality inspection, and predictive upkeep, with the mannequin appearing as a conversational layer over manufacturing information. Protection is the extra delicate deployment — intelligence evaluation, simulation help, and multilingual translation, run underneath tight safety with customized fine-tuning on defense-relevant information.
In bio and healthcare, the plans cowl literature mining, drug discovery help, and scientific documentation summarization, although no particular deployments have been named but. K2 can be headed into company instruments for doc drafting, assembly summarization, and inside data search — the unglamorous work that tends to be the place enterprise AI really earns its maintain.
Shopper purposes and multimodal derivatives
For on a regular basis customers, a light-weight variant of K2 already powers options in A., SK Telecom’s AI assistant, dealing with name summarization and computerized schedule extraction from conversations. The larger client story is the deliberate authorities service. Beneath the “AI for All” initiative, K2 is anticipated to function the spine for a free, nationwide civilian AI providing geared toward private productiveness and training — particulars of which stay fluid.
Alongside the principle mannequin, SK Telecom introduced three light-weight derivatives designed to deal with combos of textual content, speech, and pictures, with A.X K2 Mild serving as the bottom language mannequin beneath them. Probably the most fascinating of the bunch is A.X K2 Raon-Speech, a 21-billion-parameter speech basis mannequin co-developed with KRAFTON AI. It pairs a 317M-parameter speech encoder and a 96M speech codec with K2 Mild to take spoken enter and generate each textual content and artificial speech in response. The KRAFTON involvement is a helpful sign that Korea’s sovereign AI push extends past one telecom large.
Open-source ecosystem
The Hugging Face launch contains mannequin weights and configuration particulars appropriate with widespread inference frameworks, and it slots into the Ministry of Science and ICT’s analysis of home basis fashions as strategic infrastructure. The purpose is a self-sustaining Korean software program ecosystem, and open weights are an affordable strategy to seed one.
After which there’s the sensible query no person has answered. Working a 688B-parameter mannequin, even a sparse one, calls for critical compute, and the power necessities and environmental prices are undocumented. For the smaller Korean organizations this ecosystem is supposedly being constructed for, which may be the largest barrier of all. A.X K2 is a reputable technical achievement and a transparent assertion of nationwide intent — however the gaps between the announcement and the accountability are precisely the place the follow-up reporting must go.
