🆓 Free API until Jan 28th, 2026! Try on ⬆️ FriendliAI ✈️

K-EXAONE-236B-A23B-FP8

Introduction

We introduce K-EXAONE, a large-scale multilingual language model developed by LG AI Research. Built using a Mixture-of-Experts architecture, K-EXAONE features 236 billion total parameters, with 23 billion active during inference. Performance evaluations across various benchmarks demonstrate that K-EXAONE excels in reasoning, agentic capabilities, general knowledge, multilingual understanding, and long-context processing.

Key Features

• Architecture & Efficiency: Features a 236B fine-grained MoE design (23B active) optimized with Multi-Token Prediction (MTP), enabling self-speculative decoding that boosts inference throughput by approximately 1.5x.
• Long-Context Capabilities: Natively supports a 256K context window, utilizing a 3:1 hybrid attention scheme with a 128-token sliding window to significantly minimize memory usage during long-document processing.
• Multilingual Support: Covers 6 languages: Korean, English, Spanish, German, Japanese, and Vietnamese. Features a redesigned 150k vocabulary with SuperBPE, improving token efficiency by ~30%.
• Agentic Capabilities: Demonstrates superior tool-use and search capabilities via multi-agent strategies.
• Safety & Ethics: Aligned with universal human values, the model uniquely incorporates Korean cultural and historical contexts to address regional sensitivities often overlooked by other models. It demonstrates high reliability across diverse risk categories.

For more details, please refer to the technical report and GitHub.

Model Configuration

• Number of Parameters: 236B in total and 23B activated
• Number of Parameters (without embeddings): 234B
• Hidden Dimension: 6,144
• Number of Layers: 48 Main layers + 1 MTP layers
  • Hybrid Attention Pattern: 12 x (3 Sliding window attention + 1 Global attention)
• Sliding Window Attention
  • Number of Attention Heads: 64 Q-heads and 8 KV-heads
  • Head Dimension: 128 for both Q/KV
  • Sliding Window Size: 128
• Global Attention
  • Number of Attention Heads: 64 Q-heads and 8 KV-heads
  • Head Dimension: 128 for both Q/KV
  • No Rotary Positional Embedding Used (NoPE)
• Mixture of Experts:
  • Number of Experts: 128
  • Number of Activated Experts: 8
  • Number of Shared Experts: 1
  • MoE Intermediate Size: 2,048
• Vocab Size: 153,600
• Context Length: 262,144 tokens
• Knowledge Cutoff: Dec 2024 (2024/12)
• Quantization: FP8 Quantization

Evaluation Results

The following table shows the evaluation results of the quantized model compared to the original K-EXAONE model. Evaluation results of the original model against other models are available on the GitHub page or in the model card of the original model. Detailed evaluation configurations and results can be found in the technical report.

Quantization Type	AIME 2025	LiveCodeBench v6	GPQA-Diamond	KMMLU-Pro	IFBench
BF16	92.76	80.72	79.10	67.33	67.30
FP8	92.76	80.43	79.80	67.44	66.92

Requirements

Until the libraries officially support K-EXAONE, you need to install the requirements in our version with the EXAONE-MoE implementations. We will announce when these libraries are updated to support the K-EXAONE model.

Transformers

You can install the latest version of Transformers with support for EXAONE-MoE architecture from this repository. The base version of Transformers is 5.0.0rc1, so it might be helpful to check the migration guide from the Transformers library.

For quantized weights, you also need to install compressed-tensors as well. You can install it as below:

pip install "compressed-tensors>=0.13.0"

vLLM

You should install both Transformers and vLLM to use K-EXAONE model on vLLM server. You can install the latest version of vLLM with support for EXAONE-MoE architecture from this repository.

SGLang

You should install both Transformers and SGLang to use K-EXAONE model on SGLang server. You can install the latest version of SGLang with support for EXAONE-MoE architecture from this repository.

Quickstart

You can use the K-EXAONE model with the Transformers library. For better quality, you should check the usage guideline section.

Reasoning mode

For tasks that require accurate results, you can run the K-EXAONE model in reasoning mode as below.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "LGAI-EXAONE/K-EXAONE-236B-A23B-FP8"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [
    {"role": "system", "content": "You are K-EXAONE, a large language model developed by LG AI Research in South Korea, built to serve as a helpful and reliable assistant."},
    {"role": "user", "content": "Which one is bigger, 3.9 vs 3.12?"}
]
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    enable_thinking=True,   # skippable (default: True)
)

generated_ids = model.generate(
    **input_ids.to(model.device),
    max_new_tokens=16384,
    temperature=1.0,
    top_p=0.95,
    do_sample=True,
)
output_ids = generated_ids[0][input_ids['input_ids'].shape[-1]:]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

Non-reasoning mode

For tasks where latency matters more than accuracy, you can run the K-EXAONE model in non-reasoning mode as below.

messages = [
    {"role": "system", "content": "You are K-EXAONE, a large language model developed by LG AI Research in South Korea, built to serve as a helpful and reliable assistant."},
    {"role": "user", "content": "Explain how wonderful you are"}
]
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    enable_thinking=False,
)

generated_ids = model.generate(
    **input_ids.to(model.device),
    max_new_tokens=1024,
    temperature=1.0,
    top_p=0.95,
    do_sample=True,
)
output_ids = generated_ids[0][input_ids['input_ids'].shape[-1]:]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

Agentic tool use

For your AI-powered agent, you can leverage K-EXAONE’s tool calling capability. The K-EXAONE model is compatible with both OpenAI and HuggingFace tool calling specifications. The example below demonstrates tool calling using HuggingFace’s docstring-to-tool-schema utility.

Please check the example file for an example of a search agent conversation using K-EXAONE.

from transformers.utils import get_json_schema

def roll_dice(max_num: int):
    """
    Roll a dice with the number 1 to N. User can select the number N.

    Args:
        max_num: The maximum number on the dice.
    """
    return random.randint(1, max_num)

tool_schema = get_json_schema(roll_dice)
tools = [tool_schema]

messages = [
    {"role": "system", "content": "You are K-EXAONE, a large language model developed by LG AI Research in South Korea, built to serve as a helpful and reliable assistant."},
    {"role": "user", "content": "Roll a D20 twice and sum the results."}
]
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    tools=tools,
)

generated_ids = model.generate(
    **input_ids.to(model.device),
    max_new_tokens=16384,
    temperature=1.0,
    top_p=0.95,
    do_sample=True,
)
output_ids = generated_ids[0][input_ids['input_ids'].shape[-1]:]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

Usage Guideline

To achieve the expected performance, we recommend using the following configurations:

• We strongly recommend to use temperature=1.0, top_p=0.95, presence_penalty=0.0 for best performance.
• Different from EXAONE-4.0, K-EXAONE uses enable_thinking=True as default. Thus, you need to set enable_thinking=False when you want to use non-reasoning mode.

Deployment

TensorRT-LLM

TensorRT-LLM provides official support for the K-EXAONE model. Please refer to the EXAONE Documentation in the TensorRT-LLM repository for more information.

vLLM

We support the K-EXAONE model on vLLM. You need to install our fork of the vLLM library to use the K-EXAONE model. Please check the requirements section. Practically, you can serve the FP8-quantized model with a 256K context length using tensor parallel on 2 H200 GPUs.

After you install the vLLM library with an EXAONE-MoE implementation, you can run the vLLM server by following command:

vllm serve LGAI-EXAONE/K-EXAONE-236B-A23B-FP8 \
    --reasoning-parser deepseek_v3 \
    --tensor-parallel-size 2 \
    --gpu_memory_utilization 0.94 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes

An OpenAI-compatible API server will be available at http://localhost:8000/v1.

You can test the vLLM server by sending a chat completion request as below:

curl -X POST http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "LGAI-EXAONE/K-EXAONE-236B-A23B-FP8",
        "messages": [
            {"role": "user", "content": "How many r'\''s in \"strawberry\"?"}
        ],
        "max_tokens": 16384,
        "temperature": 1.0,
        "top_p": 0.95,
        "chat_template_kwargs": {"enable_thinking": true}
    }'

If you are interested in using MTP weights for speculative decoding, add according options as below.

vllm serve LGAI-EXAONE/K-EXAONE-236B-A23B-FP8 \
    --reasoning-parser deepseek_v3 \
    --tensor-parallel-size 2 \
    --gpu_memory_utilization 0.94 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes \
    --no-enable-prefix-caching \
    --speculative_config '{
        "method": "mtp", 
        "num_speculative_tokens": 2 
    }'

SGLang

We support the K-EXAONE model on SGLang. You need to install our fork of the SGLang library to use the K-EXAONE model. Please check the requirements section. Practically, you can serve the model with a 256K context length using tensor parallel on 4 H200 GPUs.

python -m sglang.launch_server \
    --model LGAI-EXAONE/K-EXAONE-236B-A23B-FP8 \
    --tp-size 4 \
    --reasoning-parser qwen3

A SGLang server will be available at http://localhost:30000.

Currently, using the OpenAI-compatible server is incompatible with the transformers>=5.0.0rc0, so you need to use SGLang native API for now. For native API, please refer to the official documentation.

Once the issue is resolved, we will update this section accordingly.

You can test the SGLang server by sending a request as below:

from transformers import AutoTokenizer
import requests

model_name = "LGAI-EXAONE/K-EXAONE-236B-A23B-FP8"
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [
    {"role": "user", "content": "How many r'\''s in \"strawberry\"?"}
]
input_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    return_tensors="pt",
)

response = requests.post(
    f"http://localhost:30000/generate",
    json={
        "text": input_text,
        "sampling_params": {
            "temperature": 1.0,
            "top_p": 0.95,
            "max_new_tokens": 16384,
        },
    },
)
print(response.json()['text'])

If you are interested in in using MTP weights for speculative decoding, add according options as below.

python -m sglang.launch_server \
    --model LGAI-EXAONE/K-EXAONE-236B-A23B-FP8 \
    --tp-size 4 \
    --reasoning-parser qwen3 \
    --speculative-algorithm EAGLE \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4

Limitation

The K-EXAONE language model has certain limitations and may occasionally generate inappropriate responses. The language model generates responses based on the output probability of tokens, and it is determined during learning from training data. While we have made every effort to exclude personal, harmful, and biased information from the training data, some problematic content may still be included, potentially leading to undesirable responses. Please note that the text generated by K-EXAONE language model does not reflect the views of LG AI Research.

• Inappropriate answers may be generated, which contain personal, harmful or other inappropriate information.
• Biased responses may be generated, which are associated with age, gender, race, and so on.
• The generated responses rely heavily on statistics from the training data, which can result in the generation of semantically or syntactically incorrect sentences.
• Since the model does not reflect the latest information, the responses may be false or contradictory.

LG AI Research strives to reduce potential risks that may arise from K-EXAONE language models. Users are not allowed to engage in any malicious activities (e.g., keying in illegal information) that may induce the creation of inappropriate outputs violating LG AI's ethical principles when using K-EXAONE language models.

License

The model is licensed under K-EXAONE AI Model License Agreement

Citation

@article{k-exaone,
  title={K-EXAONE Technical Report},
  author={{LG AI Research}},
  journal={arXiv preprint arXiv:2601.01739},
  year={2025}
}

Contact

LG AI Research Technical Support: [email protected]