The Most Openly Licensed Frontier Model Is Korean

Three frontier-scale open models shipped in the space of four days, and each came with a different definition of "open."

Kimi K3 arrived under near-MIT terms with a carve-out for Model-as-a-Service operators above $20M revenue. AMD's Instella-MoE released more of its training pipeline than anyone — every intermediate checkpoint, the data mixtures, the recipes — but licensed the weights ResearchRAIL: academic and research use only, no commercial deployment. Then on July 29, SK Telecom published A.X K2: 688 billion parameters, Apache-2.0, research and commercial, no conditions.

The most permissive licence at the frontier now belongs to a Korean telecom company.

It's a state programme, and that's the point

A.X K2 is the second model from Korea's Dopamo initiative — 독자 파운데이션 모델, "sovereign foundation model" — the government-backed effort to give Korea a frontier model it owns rather than licenses. A.X K1 came in February at 519B parameters. This is its successor, five months later.

That context explains the licence. A commercial lab prices access; a sovereignty programme is trying to establish a domestic ecosystem, and Apache-2.0 is the fastest way to get Korean manufacturers, defence contractors and hospitals building on a model that isn't American or Chinese. The permissiveness isn't generosity, it's industrial policy — the same logic behind Soofi's German industrial cloud and Mistral's positioning in France, but executed with a much looser grant than either.

Which produces an odd result: the most restrictive licence of the three came from the commercial Chinese lab, the middle one from an American chipmaker, and the freest from a state-backed national champion. Whatever intuitions people carry about which jurisdictions produce open AI, this week didn't confirm them.

Bigger, and trained on less — deliberately

The technical report contains the detail I found most interesting, and it runs against the grain of how scaling is usually discussed.

A.X K2 is larger than A.X K1 — 688B against 519B — but it was trained on fewer tokens: roughly 8.5T total, about 8.2T of that in pre-training, using what SKT describes as a smaller but higher-quality mixture. The architecture is guided by MoE scaling laws, and the stated objective is to maximize knowledge capacity within fixed hardware constraints while prioritizing inference throughput over strictly compute-optimal training.

That's a deliberate departure from Chinchilla-style compute-optimality, and the reasoning is sound once you notice who the model is for. Compute-optimal scaling minimizes training loss per unit of training compute. It says nothing about serving cost. If your deployment target is factories, defence systems and hospitals running the model continuously for years, the training run is a one-time expense and inference is the recurring one — so you buy parameters (knowledge capacity) rather than tokens, and you shape the architecture around throughput.

More parameters, less data, sparser activation. It's a serving-cost optimization wearing a scaling-law argument.

Sparse Gated Attention

The architecture centres on an attention sublayer SKT calls SGA, and it's a composite of several ideas that have been converging across labs all year:

A lightweight indexer scores and ranks key–value candidates using position scores
A selector keeps only the top-k indices
MLA (multi-head latent attention) then runs over just that selected KV cache, with compact KV and a RoPE split
A head-specific output gate, G = σ(W_g q_latent), modulates the attention output before the output projection
GatedNorm precedes the whole sublayer, doing outlier smoothing and sink control to keep low-precision FP8/FP4 computation stable

Two things stand out. First, this is sparse attention and latent attention composed together rather than treated as alternatives — you cut the number of KV entries attended to, then compress what remains. Second, the GatedNorm design exists specifically to make FP8/FP4 numerically viable, which is another inference-cost decision baked into the architecture rather than bolted on afterwards. Compare Kimi K3, which reached the same destination differently: native MXFP4 quantization-aware training from the SFT stage onward.

Everyone at the frontier is now designing for four-bit serving. They're just entering from different doors.

The benchmarks, read honestly

SKT reports thinking-mode results against four open models from the preceding six months:

Benchmark	A.X K2	Qwen3.5-397B	DeepSeek-V4 Flash	GLM-5.1	Kimi K2.6
AIME26 (math)	97.1	92.5	96.7	95.4	95.4
Apex (math)	45.8	13.5	28.1	10.4	20.8
KMMLU-Pro (Korean)	80.5	78.4	78.8	76.5	73.4
CLIcK (Korean)	91.6	88.4	89.6	88.8	80.9
τ²-Bench Telecom (agentic)	98.0	95.6	95.0	97.7	95.9

It leads on all five. But the five are not equally informative, and it's worth separating them.

Two are home turf. KMMLU-Pro and CLIcK are Korean-language benchmarks, and a Korean-language model built by a Korean company leading them is the expected outcome, not evidence of general capability. The margins are also modest — one to two points over DeepSeek-V4 Flash.

One is home turf in a subtler way. τ²-Bench Telecom is an agentic benchmark in the telecommunications domain, published by a telecommunications company. 98.0 is a strong number, and it's also 0.3 points above GLM-5.1 on a saturated benchmark in SKT's own industry. That's the kind of result to note rather than lean on.

One is genuinely striking. On Apex, A.X K2 scores 45.8 where the next best is DeepSeek-V4 Flash at 28.1 — roughly 1.6× the field, with GLM-5.1 and Qwen3.5 in the low teens. A gap that size on a hard math benchmark isn't a tuning artifact or a home-turf effect; either something in the training mixture or the reasoning post-training is doing real work, or the benchmark interacts badly with the other models' harnesses. It's the single result most worth independent replication.

And AIME26 at 97.1 is a lead of 0.4 points over DeepSeek-V4 Flash. At that level the benchmark is effectively saturated and the ordering is close to noise.

All figures are SKT's own, in thinking mode, on a chart in their own report — the same caveat that applies to every vendor benchmark table, including Celeris's and Microsoft's.

Where this actually lands

The deployment plan is the least glamorous and most telling part: manufacturing, defence and bio, with named partners including KG Steel and Connec. Not a chat assistant. Not a coding product. Specialized industrial agents for domestic heavy industry.

That's what a sovereignty model is for, and it clarifies why the Apache-2.0 licence and the inference-throughput architecture belong to the same decision. A Korean steelmaker deploying an agent on its own plant floor needs three things: no licensing conditions that complicate commercial use, serving costs low enough to run continuously, and no dependency on a foreign API. A.X K2 is engineered against exactly that specification, and the frontier benchmark scores are almost a side effect.

It also punctures a tidy assumption I've made elsewhere — that openness at the frontier narrows as models get larger, with licences quietly selecting for the few organizations able to run them. A.X K2 is 688B parameters and freer than a 16B research release from AMD. The licence gradient doesn't track scale. It tracks what the releasing organization wants to happen next — and a state building an industrial base wants adoption more than it wants rent.