A.X K2
Model Summary
A.X K2 is a large-scale Mixture-of-Experts (MoE) language model trained from scratch as a high-performance, agentic foundation model, and the successor to A.X K1. The model contains 688 billion total parameters, with 33 billion active parameters, delivering strong reasoning and instruction-following performance while maintaining practical inference efficiency.
Through a Think-Fusion training recipe, a single unified model supports both a thinking mode for complex problem solving and a non-thinking mode for concise, low-latency responses, allowing the user to trade quality for cost on a per-request basis.
A.X K2 is developed as part of the Korean government's Sovereign AI foundation model project, aiming to build a frontier-scale model with deep understanding of the Korean language and culture.
A.X K2 NVFP4
A.X K2 NVFP4 is an NVFP4-quantized version of A.X K2, a 688B-parameter Mixture-of-Experts language model developed by SK Telecom.
This checkpoint applies NVFP4 W4A4 quantization to the routed experts, while keeping the remaining modules in FP8 or BF16. It retains performance comparable to the FP8 checkpoint while reducing the model memory footprint by approximately half.
The model can be served on a single node with 4 NVIDIA B200 GPUs using the A.X K2–enabled SKT-AI vLLM fork.
Quantization
| Module | Precision |
|---|---|
| Routed experts | NVFP4 W4A4 |
| Attention, router, shared expert, and dense layer | FP8 |
| Embedding and LM head | BF16 |
Usage
Install the A.X K2–enabled vLLM fork:
git clone -b axk2-v0.23.0 https://github.com/SKT-AI/vllm.git
cd vllm
pip install -e .
Enable the FlashInfer NVFP4 MoE kernels:
export VLLM_USE_FLASHINFER_MOE_FP4=1
export FLASHINFER_DISABLE_VERSION_CHECK=1
Serve the model:
vllm serve skt/A.X-K2-NVFP4 \
--served-model-name A.X-K2-NVFP4 \
--host 0.0.0.0 \
--port 8000 \
--tensor-parallel-size 4 \
--enable-expert-parallel \
--attention-backend FLASHINFER_MLA_SPARSE \
--reasoning-parser deepseek_v3
The A.X K2 implementation is registered directly in the SKT-AI vLLM branch, so --trust-remote-code is not required.
To enable automatic tool calling, add:
--enable-auto-tool-choice \
--tool-call-parser hermes
Thinking Mode
Thinking mode can be controlled per request through chat_template_kwargs.
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "A.X-K2-NVFP4",
"messages": [
{
"role": "user",
"content": "Explain how speculative decoding works."
}
],
"max_tokens": 1024,
"chat_template_kwargs": {
"enable_thinking": true
}
}'
Set enable_thinking to false for a direct non-thinking response.
Long Context
The checkpoint includes the YaRN configuration required for context lengths of up to 256K tokens. No separate RoPE override is required when using the provided config.json.
The serving context can be reduced with --max-model-len when longer inputs are not required.
Needle-in-a-Haystack (NIAH)
NIAH probes exact fact retrieval by inserting a target fact (the "needle") at varying depths within a long context and asking the model to recover it. Under zero-shot YaRN scaling (scaling factors of 1, 2, and 4 for 128K, 256K, and 512K, respectively), A.X K2 attains a perfect retrieval score at every context length and needle depth—and does so even after NVFP4 (experts-only W4A4) quantization. The heatmaps below show the 256K and 512K results.
License
A.X K2 is released under the Apache License 2.0.
Citation
@techreport{axk2-2026,
title = {A.X K2 Technical Report},
author = {SK Telecom},
year = {2026},
institution = {SK Telecom},
url = {https://github.com/SKT-AI/A.X-K2/blob/main/A_X_K2_Tech_Report.pdf}
}