News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
A.X K1

A.X K1

Sparse MoE model with controllable reasoning depth

A.X K1 is a large language model built by SK Telecom. Released in 2026, it uses a sparse mixture of experts design to balance capacity and speed. The system allows users to toggle between deep reasoning steps and fast direct answers.

Share on WhatsApp

  • 519 billion total parameters with 33 billion active per token
  • Supports user-controlled switch between deep reasoning and fast response
  • Context window extends to 131,072 tokens
  • Optimized tokenizer for five languages including Korean and code
  • Uses multi-token prediction during training for stability

What it is

A.X K1 is a large-scale language model developed by SK Telecom. It features 519 billion total parameters but activates only 33 billion for each token processed. This sparse design aims to provide high reasoning capacity while keeping inference costs manageable. The model supports a context window of 131,072 tokens and handles text in multiple languages. It was trained from scratch rather than building on an existing pretrained base.

How it came about

The model was announced in 2026 via a technical report authored by a large team at SK Telecom. It represents a standalone training effort, as the developers state it was not initialized from any prior model. The release includes integration support for common inference frameworks such as vLLM and SGLang, indicating a focus on practical deployment from the outset.

Inside the model

The architecture is a decoder-only transformer using a mixture of experts approach. It contains 192 experts plus one shared expert, with 61 layers in total. A key structural choice is the application of RMSNorm both before and after the MLP blocks, which the team says improves stability in sparse settings. Training employed a multi-token prediction objective to help stabilize the process. The tokenizer uses byte pair encoding and is optimized for English, Korean, Chinese, Japanese, and Spanish, with special attention to code patterns.

What it can do

The model offers a hybrid reasoning mode. Users can select a think mode for complex multi-step problems or a non-think mode for faster, concise replies. In benchmarks, it performs strongly on Korean language tasks such as KMMLU and math tests like AIME25. It also shows competitive results in code generation benchmarks including HumanEval+ and LiveCodeBench. Performance varies between the two reasoning modes, with the thinking mode generally yielding higher scores on complex logic tasks.

How to use it

Weights are available openly through Hugging Face. The developers provide example scripts for running the model with Transformers, vLLM, and SGLang. These integrations support multi-node and tensor-parallel configurations. For lightweight or quantized versions suited to business use, interested parties must contact SK Telecom directly.

Safety and testing

The documentation notes that the model may produce incorrect or misleading information due to its stochastic nature. It warns that the reasoning steps shown in think mode should not be viewed as a faithful record of the internal decision process. Performance consistency depends on data coverage across different domains and languages.

Questions people ask about A.X K1

How does the mixture of experts design work in this model?

The model contains 192 experts but only activates eight of them plus one shared expert for each token. This keeps the active parameter count at 33 billion, allowing high total capacity without the full computational cost of a dense 519 billion parameter model.

Can I choose how much the model reasons before answering?

Yes. The model features a hybrid reasoning control. You can enable think mode for detailed multi-step reasoning on complex tasks, or disable it for concise, low-latency responses. Both modes operate within the same unified model structure.

Which languages does A.X K1 support best?

The tokenizer is optimized for English, Korean, Chinese, Japanese, and Spanish. Benchmark results show particularly strong performance on Korean language evaluations such as KMMLU and KoBALT, reflecting a specific emphasis on these regions.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on A.X K1 yet, so the facts on this page come from the sources listed above: huggingface.co/skt/A.X-K1. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.