SK Telecom released A.X K2, a large Mixture-of-Experts language model it trained from scratch, as Apache-2.0 open weights on Hugging Face on July 29, 2026. The model is 688 billion total parameters with 33 billion active per token, and it posts 80.5 on KMMLU-Pro and 91.6 on CLIcK, the highest scores among the seven competing models tabulated in its own card. On BrowseComp, the web-search agentic task, it scores 9.3, far behind GLM-5.1 at 29.1 and Qwen3.5 at 26.9. ASAP works from SK Telecom's official model card as the primary source to map where this model arrived and where it stalled.

Korean and telecom lead, code and search trail

A.X K2's benchmark table is split cleanly by domain, with Korean and telecom tasks at the top and code, knowledge and web search well down the field. In thinking mode, two of the three Korean tasks lead the field, KMMLU-Pro at 80.5 and CLIcK at 91.6, and the telecom agentic task τ²-Bench Telecom also comes first at 98.0. In mathematics, AIME26 at 97.1 is the top score, and on the harder Apex set the model reaches 45.8 against 28.1 for the runner-up, DeepSeek-V4 Flash.

The other side is equally clear. LiveCodeBench v6 at 84.0 sits behind DeepSeek-V4 Flash at 89.4 and Kimi-K2.6 at 86.5, and SciCode at 41.0 trails Kimi-K2.6 at 53.5 by 12.5 points. Knowledge tasks place mid-pack, with Humanity's Last Exam at 27.8 and GPQA Diamond at 85.6. The widest gap is BrowseComp, where 9.3 is roughly a third of GLM-5.1's 29.1.

Even within Korean the result is not uniform. KMMLU-Pro and CLIcK lead, but KoBALT, which probes linguistic reasoning, comes in at 73.0 against DeepSeek-V4 Flash's 75.3. "First place in Korean" describes two of three tasks, not all three.

Why the gain over the previous model varies so widely by task

Averaging the point differences across the 14 tasks where A.X K1 also has a score gives a gain of 18.2 points (calculated by ASAP directly from the model card table). That average carries little information on its own, because the per-task gains are extremely dispersed.

The largest movement is Apex, from 1.0 to 45.8, a 44.8-point gain. Apex-shortlist follows, from 50.8 to 88.6, then AA-LCR from 26.0 to 66.0 and KoBALT from 51.0 to 73.0. At the other end, AIME26 moves only from 91.3 to 97.1, and CLIcK from 85.3 to 91.6.

That distribution says something about headroom. On tasks where K1 already cleared 85, such as AIME26 and CLIcK, little room remained, and nearly all of the gain came from tasks K1 essentially could not do. K2's improvement therefore reads as entry into a previously closed difficulty band rather than uniform across-the-board progress. A K1 score of 1.0 on Apex marks that task as a wall for the prior generation, and 45.8 marks the wall as partially breached. It is also still a score below half marks.

Running 688 billion parameters at 33 billion

A.X K2 is a Mixture-of-Experts model with 256 routed experts plus 1 shared expert, activating 8 experts per token. Of its 61 layers, 1 is dense and 60 are MoE. It uses 64 attention heads, a hidden size of 7,168 and a vocabulary of 163,840. Relative to A.X K1, only the routed expert count changed, rising from 192 to 256, while the 33B active parameter count stayed fixed.

Long context runs on Sparse Gated Attention. A lightweight sparse-attention indexer sits on top of Multi-head Latent Attention, selecting the top 2,048 tokens per query, while head-specific gated attention suppresses attention sinks. Context length is 262,144 tokens, trained natively to 128K and extended to 256K with YaRN scaling. RULER averages 94.6, declining gradually with length: 97.5 at 4K, 92.7 at 128K and 86.6 at 256K.

The most consequential training choice is native FP8 (MXFP8, E4M3) for both forward and backward passes from the start. The released checkpoint ships in block-scaled FP8, so there is no separate post-hoc quantization step. Pre-training ran on roughly 8.2 trillion tokens, filtered down from a 16.2-trillion-token candidate pool and arranged as a three-stage curriculum of 6.4T general knowledge, 1.4T high-quality reasoning and 0.36T long-context data. Post-training added roughly 60.6B tokens of supervised fine-tuning plus reinforcement learning over about 100K prompts.

What "sovereign AI" actually covers here

The model card states that A.X K2 was developed as part of the Korean government's Sovereign AI foundation model project. That phrase should be read alongside the data composition the same card publishes. The designed language mix of the first training stage is 72.7% English, 15.4% Korean and 8.3% code, with Japanese, Spanish and Chinese at roughly 1% each, and later stages shift further toward English. Korean web text comes from an in-house crawl of approximately 1.37 trillion tokens plus an in-house PDF parsing pipeline and synthetic data, none of which is released.

The largest single contributor to the English web corpus is NVIDIA's Nemotron-CC-v2.1, with code from Nemotron-Pretraining-Code-v2 and multilingual coverage from FineWeb2. The sovereign element of this model therefore lies in weight ownership and training control rather than in data independence. What is substantive is the 1.37-trillion-token in-house Korean corpus that drives its Korean performance, plus Apache-2.0 open weights as the distribution term. Most of the training data comes from public English corpora.

This distinction matters because in Korean policy discussion "sovereign AI" often expands to mean a model free of foreign data. What A.X K2 secured is different: the ability to own, modify and redistribute weights domestically without depending on an external API, and Korean-language competence lifted by proprietary data. KMMLU-Pro at 80.5 and CLIcK at 91.6 are the benchmark evidence that the investment returned. Presenting the model as an instance of full data independence contradicts its own card.

The 800 GiB threshold

Open weights and runnable weights are different claims, and A.X K2 is a clear case where the gap between them decides who can actually deploy the model. The FP8 checkpoint occupies roughly 647 GiB on disk and a little more once loaded, and the card advises planning for at least 800 GiB of aggregate GPU memory once the KV cache and activations are included. Validation used four B300 cards at 275 GiB each; twelve or more 80 GB cards would be required instead. On hardware below compute capability 8.9, or with no GPU visible, Transformers dequantizes to bfloat16 automatically, which needs roughly 1.4 TB for the full model.

Mapped onto the Korean market, that requirement draws a sharp line. For large enterprises already operating GPU clusters and for institutions inside national infrastructure programs, A.X K2 is an immediately deployable asset. For a mid-sized company or startup with one or two GPU servers, on-premise operation of a 688B model remains out of reach. SK Telecom closing the card with a contact address for lightweight versions signals awareness of that gap.

Put differently, the practical effect of the Apache-2.0 license is not "anyone can run it" but "organizations that can run it face no restrictions." Commercial use and derivative models carry no additional conditions, so building domain-specific models on this base can begin without license negotiation. The real impact will be decided by the smaller models derived from K2 rather than by K2 itself.

The limits the card writes down itself

One notable feature of this release is that SK Telecom documented the weaknesses in the card. On agentic capability it writes that the model trails the strongest compared models on BrowseComp and attributes this to limited agentic reinforcement learning during post-training, adding a recommendation to evaluate long-horizon tool-use workloads before relying on them. That is not benchmark copy written as marketing.

On bias, the card records that bias is one of nine dimensions judged during safety RL but that no dedicated quantitative bias or fairness evaluation was run. It also states that reported benchmarks are predominantly English and Korean and largely single-turn. Multi-turn behavior and quality in languages other than Korean and English cannot be confirmed from this document.

The point requiring the most care is that the entire comparison table is self-reported. The card does not establish the conditions under which the seven competing models were reproduced, nor how thinking-mode budgets and prompt settings were matched. It links a separate technical report PDF, so any adoption decision resting on these numbers should verify the evaluation protocol there. Apache-2.0 open weights do mitigate the problem, because third-party reproduction is possible.

What to watch over the next six months

A.X K2 will be judged on three follow-on indicators rather than on its benchmark rank. The first is agentic improvement. A BrowseComp score of 9.3 means the model is not yet competitive on long tool-using tasks, and the question is how much of the agentic RL that SK Telecom named as the cause lands in the next version. For a model positioned as an agentic foundation model, that gap goes to its identity.

The second is the lightweight lineage. Few Korean organizations can run a 688B model as released, so actual adoption depends on when derivative small models arrive and under what license. The third is third-party reproduction. The chief benefit of an open-weight release is that outside researchers can redraw the same table, and whether KMMLU-Pro 80.5 and τ²-Bench Telecom 98.0 survive independent evaluation is what will settle confidence in this model.

Source: ASAP analysis based on SK Telecom's official model card skt/A.X-K2 (Hugging Face, Apache-2.0) and the linked A.X K2 Technical Report