A.X K2

SK Telecom's 688B mixture of experts with 33B active, native FP8 weights and a 256K context. Apache 2.0, strongest at math and Korean.

In short. A.X K2 is a very large AI model from the South Korean telecom company SK Telecom. It is free to use commercially under Apache 2.0, it is especially good at math and Korean-language questions, and its step-by-step thinking can be switched on or off. It reads text only and needs a big multi-GPU server.

Summary

SK Telecom released A.X K2 on Hugging Face on 29 July 2026, as its second-phase model for South Korea's government-backed sovereign AI foundation model project. It is a text-only mixture-of-experts model with 688 billion parameters and 33 billion active per token: 61 layers (1 dense and 60 mixture-of-experts layers), 256 routed experts with 8 active plus one shared expert, and Multi-head Latent Attention with a sparse indexer that reads 2,048 positions per query. It was trained from scratch, natively in FP8, on roughly 8 trillion tokens, with a native 128K context and 256K reached through YaRN scaling. One checkpoint offers a thinking mode and a non-thinking mode through SK Telecom's Think-Fusion recipe. The tokenizer and training data target English, Korean, Chinese, Japanese and Spanish.

In SK Telecom's own table, run in thinking mode against seven other models, A.X K2 has the best score on AIME26 (97.1), Apex, Apex-shortlist, KMMLU-Pro (80.5), CLIcK (91.6) and τ²-Bench Telecom (98.0). It trails the strongest models on agentic and coding tests: it scores 36.0 on Terminal Bench v2.1 against 61.8 for DeepSeek-V4 Flash and GLM-5.1, and the card itself calls its agentic performance moderate. SK Telecom reports an average gain of 32.2 points over A.X K1 across 14 benchmarks. These are SK Telecom's results, and the rows marked with an asterisk were run by Artificial Analysis.

The weights are under Apache 2.0. The checkpoint is block-scaled FP8 at about 694 GB, with an NVFP4 build of about 398 GB, and the card says to plan for at least 800 GiB of aggregate GPU memory including the KV cache; its reference runs used four 275 GB GPUs. Transformers supports the model natively but only from a source install, because the support was merged after v5.14.1, and vLLM needs SK Telecom's fork based on v0.23.0. The card points to vLLM rather than Transformers for long context and serving.

Good for

  • Math and Korean-language work, where SK Telecom's own table shows its best results (AIME26 at 97.1 and KMMLU-Pro at 80.5)
  • Long-context tasks up to 256K tokens, where the card reports a RULER average of 94.6
  • Teams that want a self-hosted, Apache 2.0 model developed in Korea, with a thinking mode that can be turned off for cheaper replies
  • Serving at low precision, since the weights are native FP8 and the report says NVFP4 stays within one point of FP8 accuracy

Watch out for

  • Agentic performance is moderate, and the card says it trails the strongest compared models on Terminal Bench and BrowseComp
  • It is text-only, and only five languages are targeted (English, Korean, Chinese, Japanese and Spanish)
  • The FP8 weights are about 694 GB and the card recommends at least 800 GiB of GPU memory, so it needs a multi-GPU node
  • vLLM support comes from SK Telecom's fork and Transformers needs a source install, and flash attention is not supported under Transformers
  • Benchmark numbers are SK Telecom's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the A.X K2 NVFP4 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
A.X K2 (FP8) FP8 688B 33B active 1.58 TB 791 GB 475 GB Big-memory desktop
A.X K2 NVFP4 NVFP4 688B 33B active 1.58 TB 791 GB 438 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is A.X K2?
A.X K2 is a very large AI model from the South Korean telecom company SK Telecom. It is free to use commercially under Apache 2.0, it is especially good at math and Korean-language questions, and its step-by-step thinking can be switched on or off. It reads text only and needs a big multi-GPU server.
Can I use A.X K2 commercially?
Yes. A.X K2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is A.X K2 open source?
A.X K2 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run A.X K2 on my own computer?
Only on a machine with a lot of memory. The smallest variant, A.X K2 NVFP4, needs about 438 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was A.X K2 released?
A.X K2 was released by SK Telecom on July 29, 2026.
How long a context window does A.X K2 have?
A.X K2 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
K-EXAONE 2.0 LG AI Research 750B 37B active 256K Apache 2.0 (Permissive) 518 GB GPU server
Solar Open 2 Upstage 250B 15B active 1M Upstage Solar (Conditional) 157 GB Big-memory desktop
Motif 3 Motif Technologies 314B 13.2B active 256K MIT (Permissive) 206 GB Big-memory desktop
DeepSeek V4 DeepSeek 284B – 1.6T 49B active 1M MIT (Permissive) 176 GB Big-memory desktop