DEV.co
Open-Source LLM · skt

A.X-K1

A.X K1 is a 519-billion-parameter sparse Mixture-of-Experts language model from SKT that activates only 33 billion parameters per token. It supports both reasoning-intensive and fast inference modes, handles five languages (English, Korean, Chinese, Japanese, Spanish) plus code, and is available under Apache 2.0 without gating. The model was trained from scratch and shows competitive performance on knowledge, math, code, and instruction-following benchmarks, though it underperforms on some long-context tasks.

Source: HuggingFace — huggingface.co/skt/A.X-K1
519B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
33.3k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerskt
Parameters519B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads33.3k
Likes286
Last updated2026-06-29
Sourceskt/A.X-K1

What A.X-K1 is

A.X K1 is a decoder-only Transformer with 192 total experts (1 shared, 191 sparse) and 61 layers. It uses Multi-Latent Attention (MLA), RMSNorm before and after MLP blocks, a 163,840-token vocabulary optimized for multilingual and code data, and supports 131,072-token context length. Training employed multi-token prediction (MTP) as an auxiliary objective. The sparse MoE design allows scaling capacity without proportional compute increase. Model card indicates vLLM and SGLang integrations are available (via non-upstream branches); standard Hugging Face Transformers support is provided.

Quickstart

Run A.X-K1 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="skt/A.X-K1")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Korean and Multilingual AI Applications

Strong performance on Korean benchmarks (KMMLU, KoBALT, AIME25-ko) makes it suitable for Korean-language customer support, content generation, and enterprise applications. Tokenizer optimized for multiple languages reduces token overhead.

Reasoning-Heavy Problem Solving with Latency Flexibility

Think/Non-Think modes allow a single deployment to handle both complex multi-step reasoning (math, planning) and fast-response use cases (chat, real-time agents). Think mode shows strong AIME25 performance (89.8%).

Large-Scale Self-Hosted Inference

Sparse MoE design and 33B active parameters enable efficient multi-GPU/multi-node inference with expert parallelism. SGLang and vLLM integrations (via provided branches) support tensor parallelism across 32+ GPUs in tested configurations.

Running & fine-tuning it

Estimated 208–280 GB GPU VRAM for full-precision inference (519B params ÷ sparse activation ratio ~1.9–2.5x overhead for routing, activation maps). Tested configuration used TP=32 across 4 nodes (128 GPUs, likely H100/A100 class). For a single 8xH100 setup: requires bfloat16 quantization or expert offloading. Exact VRAM per-GPU depends on batch size, context length (tested to 131k), and quantization; model card does not specify minimum or recommended specs.

Unknown. Model card provides no guidance on LoRA, QLoRA, or instruction-tuning feasibility for sparse MoE. Standard fine-tuning on MoE models introduces challenges (expert load balancing, gradient sparsity). Contact vendor ([email protected]) for guidance; quantized/lightweight variants (potentially easier to fine-tune) are vendor-gated.

When to avoid it — and what to weigh

  • Strict Long-Context Retrieval Requirements — Performance on long-context benchmarks (AA-LCR: 36%, Humanity's Last Exam: 8.6%) significantly trails competitors (DeepSeek-V3.1, GLM-4.6). Not recommended if needle-in-haystack or long-document QA is mission-critical.
  • Need for Guaranteed Factuality or Reasoning Transparency — Model card explicitly states outputs may be 'incorrect or misleading' and that Think-mode reasoning should not be interpreted as a 'faithful representation' of internal decision processes. Unsuitable for high-stakes applications requiring explainability.
  • Single-GPU or Edge Deployment — 519B total parameters with sparse activation still requires substantial GPU memory and distributed setup. Quantized/lightweight versions are only available via direct contact to vendor ([email protected]), limiting off-the-shelf edge feasibility.
  • Production Deployment Without Vendor Support — vLLM and SGLang integrations are on non-upstream branches (fort726 forks) and noted as 'initial' with evolving parameters. No guaranteed LTS or production SLA documented.

License & commercial use

Apache-2.0. Permissive OSI license allowing modification, distribution, and private use with no copyleft obligations. Model is ungated (gated: false).

Apache-2.0 is a permissive license generally compatible with commercial use. However, the model card states that lightweight and quantized variants are available 'for business applications or scalable deployment' only via direct vendor contact ([email protected]), implying potential commercial licensing terms or SLAs beyond the base open model. Full-parameter open model may be commercially usable as-is, but production deployment should confirm with vendor whether commercial support, indemnification, or separate terms apply.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitGood
Assessment confidenceHigh
Security considerations

No security audit, adversarial robustness testing, or red-teaming results provided. Model trained on large-scale data (composition unknown); potential for embedded biases or unvetted content echoes. Think-mode outputs should not be assumed to reflect true reasoning; possible to misinterpret confidence in incorrect answers. No data provenance, filtering, or deduplication methodology disclosed. Users deploying at scale should conduct internal adversarial testing and implement guardrails.

Alternatives to consider

DeepSeek-V3.1 (685B-37B active)

Larger total and active parameters; stronger on long-context (AA-LCR: 53.3 vs. A.X K1: 36%) and GPQA (77.9 vs. 74%), but heavier inference cost. Model card does not disclose its license/commercial terms clearly.

GLM-4.6 (357B-32B active)

Smaller total parameter count (easier deployment); competitive on MMLU-Pro, IFEval, code; mixed performance vs. A.X K1 by benchmark. Korean support and license/availability terms not specified here.

Llama 3.1 or Llama 3 (70B–405B, dense)

Fully open-source, extensive ecosystem support (quantization, fine-tuning), broader deployment simplicity. Trade-off: dense models, higher inference cost per token; no sparse-MoE efficiency gains or Think-mode reasoning control.

Software development agency

Ship A.X-K1 with senior software developers

Get expert guidance on infrastructure sizing, quantization options, and production integration. Our team can help you configure tensor parallelism, validate Think/Non-Think modes for your use case, and connect you with vendor support for business-grade variants.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

A.X-K1 FAQ

Can I use A.X K1 for commercial products under Apache-2.0?
Apache-2.0 is a permissive license that generally permits commercial use. However, the model card notes that business-grade quantized and lightweight variants are only available via vendor contact ([email protected]). For production deployment, confirm licensing terms with the vendor to ensure compliance and avoid ambiguity around support or indemnification.
What GPU setup do I need to run A.X K1 locally?
At least 8×H100 or A100 (80 GB each, ~640 GB total) for bfloat16 inference with moderate batch size and context length. The tested evaluation used 4 nodes with 32 GPUs (TP=32) and tensor parallelism. Single-GPU deployment is not practical without expert offloading or aggressive quantization. Contact vendor for quantized variant options.
How does Think mode differ from Non-Think mode at inference time?
Think mode generates explicit multi-step reasoning before the answer, improving performance on complex tasks (e.g., AIME25: 89.8%). Non-Think mode skips reasoning and generates direct responses for low-latency use cases. Both modes use the same unified model; the choice is made at inference time via prompt/token control, not a separate model.
What languages does A.X K1 support?
English, Korean, Chinese, Mandarin, Japanese, and Spanish. The tokenizer is optimized for these languages plus code. Performance varies by language and benchmark; Korean benchmarks show strong results. Coverage outside these five languages is not documented.

Custom software development services

Need help beyond evaluating A.X-K1? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy A.X K1?

Get expert guidance on infrastructure sizing, quantization options, and production integration. Our team can help you configure tensor parallelism, validate Think/Non-Think modes for your use case, and connect you with vendor support for business-grade variants.