Frontier-scale open weights almost never arrive on a schedule. These did. SK Telecom published A.X K2 — 688 billion total parameters, roughly 33 billion active per token — to Hugging Face on July 29. Two days later LG AI Research posted K-EXAONE 2.0 at 750 billion parameters, about 37 billion active. Same mixture-of-experts architecture. Same 262,144-token context window. Same Apache 2.0 license, which lets any company download, modify and ship them commercially with no royalty and no obligation to publish changes.
The 48-hour gap is not a coincidence, and neither is the licensing. Both models are entries in South Korea’s Sovereign AI Foundation Model project, a government contest run by the Ministry of Science and ICT that eliminates teams round by round. Evaluators convene August 8–11 to cut the field from four to three. A further cut by year-end leaves two, and those two anchor the technical backbone of a national AI assistant the government intends to offer free to all 51 million residents. Openness is not a philosophy here. It is a submission requirement.
The contest has already proved it bites. Five consortia started in August 2025; the plan was to drop one after the first evaluation. On January 15 the ministry dropped two. Naver Cloud — the presumed frontrunner, with HyperCLOVA X behind it and the deepest Korean-language data in the country — was cut because its submission used frozen encoder weights from Alibaba’s Qwen for video and audio. Used as-is, never updated. That failed the requirement for full in-house training. NC AI went too. A startup consortium led by Motif Technologies with KAIST and Trillion Labs was added in February to refill the bracket.
Our take: Everyone reads this as a Korea story. It is a distribution story. Two models at genuine frontier scale just landed under the most permissive license in the business, and the reason is that a procurement panel scores openness — not that anyone decided sharing was good strategy. That is a repeatable mechanism, and it is the only one that has reliably produced open frontier weights outside of China. Governments buying models is going to keep pushing weights into the commons faster than any manifesto ever has. Watch what Japan, India and the EU write into their own tenders next.
One model card says the quiet part
The most useful number in either release is a bad one. On BrowseComp — a benchmark for answering hard questions through open-ended web research — A.X K2 scored 9.3, last in SK Telecom’s own comparison set, against 29.1 for GLM-5.1 and 26.9 for Qwen3.5. SK Telecom published it anyway, attributed it to “limited agentic RL during post-training,” and told users to evaluate long-horizon tool-use workloads before relying on the model.
That is a lab telling you where its model breaks before you find out in production. It is also a precise map of where the training budget went: A.X K2 hit 97.1 on AIME26 and solved 35 of 42 International Mathematical Olympiad 2025 problems, above the gold-medal threshold. Structured reasoning, funded. Open-ended agentic search, not. If you were planning to drop it into an agent loop, that one line saves you a quarter.
What each one is actually for
K-EXAONE 2.0 leads on long documents: 94.4 on OpenAI-MRCR, ahead of Qwen3.5 at 93.0 and DeepSeek V4 Pro Max at 92.9, and 89.6 on Ko-LongBench against 83.6 for GLM-5.1. Across LG’s 24-benchmark suite it averaged 70.1 versus 63.3 for its predecessor. It got there by “upcycling” — expanding LG’s own 236-billion-parameter Phase 1 model rather than training from scratch. That is a live question for evaluators who just cut a team over weight provenance.
A.X K2 went the other way: trained natively in FP8 across roughly 8.2 trillion tokens, and shipped in block-scaled FP8 so it deploys at lower memory cost with no separate quantization step. Neither is casual to run — A.X K2’s card references a four-GPU setup with around 656GB of aggregate memory.
What to watch
- August 8–11. The second-phase panel cuts four teams to three. Two hundred citizen evaluators, applications closing August 4, feed usability results into it.
- Whether upcycling counts as sovereign. The originality standard that eliminated Naver Cloud now judges LG’s decision to build on its own prior weights. Accepted so far.
- Upstage. The startup in the field published full training logs for independent verification and claims better Korean token efficiency than either giant — 4.41 bytes per token against 4.12 for K-EXAONE. That is a direct inference-cost edge.
- Who copies the mechanism. Two frontier-scale Apache 2.0 releases in one week came out of a procurement scorecard, not a manifesto.