Hugging Face's logo

Transformers documentation

EXAONE MoE

You are viewing main version, which requires installation from source. If you'd like regular pip install, checkout the latest stable version (v5.0.0).
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

This model was released on 2025-12-31 and added to Hugging Face Transformers on 2026-02-02.

EXAONE MoE

Overview

K-EXAONE model is a large-scale multilingual language model developed by LG AI Research. Built using a Mixture-of-Experts architecture named EXAONE-MoE, 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.

All model weights including quantized version are available at Huggingface Collections.

Model Details

Model Configuration of K-EXAONE

  • 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)

Usage Tips

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"

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,
)
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,
)
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,
)
output_ids = generated_ids[0][input_ids['input_ids'].shape[-1]:]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

ExaoneMoeConfig

class transformers.ExaoneMoeConfig

< >

( vocab_size = 102400 hidden_size = 4096 intermediate_size = 16384 num_hidden_layers = 32 num_attention_heads = 32 num_key_value_heads = 32 hidden_act = 'silu' max_position_embeddings = 2048 initializer_range = 0.02 rms_norm_eps = 1e-05 use_cache = True bos_token_id = 1 eos_token_id = 53 pad_token_id = 0 tie_word_embeddings = False rope_parameters = None attention_dropout = 0.0 sliding_window = 4096 sliding_window_pattern = 4 layer_types = None mlp_layer_types = None first_k_dense_replace = 1 moe_intermediate_size = 1024 num_experts = 64 num_experts_per_tok = 8 num_shared_experts = 1 norm_topk_prob = True routed_scaling_factor = 2.5 n_group = 1 topk_group = 1 **kwargs )

Parameters

  • vocab_size (int, optional, defaults to 102400) — Vocabulary size of the EXAONE MoE model. Defines the number of different tokens that can be represented by the inputs_ids passed when calling [ExaoneMoeModel].
  • hidden_size (int, optional, defaults to 4096) — Dimension of the hidden representations.
  • intermediate_size (int, optional, defaults to 16384) — Dimensionality of the MLP representations.
  • num_hidden_layers (int, optional, defaults to 32) — Number of hidden layers in the Transformer encoder.
  • num_attention_heads (int, optional, defaults to 32) — Number of attention heads for each attention layer in the Transformer decoder.
  • num_key_value_heads (int, optional, defaults to 32) — This is the number of key_value heads that should be used to implement Grouped Query Attention. If num_key_value_heads=num_attention_heads, the model will use Multi Head Attention (MHA), if num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details checkout this paper. If it is not specified, will default to num_attention_heads*.
  • hidden_act (str or function, optional, defaults to “silu”) — The non-linear activation function (function or string) in the decoder.
  • max_position_embeddings (int, optional, defaults to 2048) — The maximum sequence length that this model might ever be used with. Typically set this to something large just in case (e.g., 32768 for EXAONE 3.5).
  • initializer_range (float, optional, defaults to 0.02) — The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
  • rms_norm_eps (float, optional, defaults to 1e-05) — The epsilon used by the layer normalization layers.
  • use_cache (bool, optional, defaults to *True) -- Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if config.is_decoder=True`.
  • bos_token_id (int, optional, defaults to 1) — Beginning of stream token id.
  • eos_token_id (int, optional, defaults to 53) — End of stream token id.
  • pad_token_id (int, optional, defaults to 0) — Padding token id.
  • tie_word_embeddings (bool, optional, defaults to False) — Whether to tie weight embeddings
  • rope_parameters (RopeParameters, optional) — Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain a value for rope_theta and optionally parameters used for scaling in case you want to use RoPE with longer max_position_embeddings.
  • attention_dropout (float, optional, defaults to 0.0) — The dropout ratio for the attention probabilities.
  • sliding_window (int, optional, defaults to 4096) — The size of the sliding window for the sliding window attention.
  • sliding_window_pattern (str, optional, defaults to 4) — The pattern to use for sliding window attention. Can be one of:
    • None: No sliding window attention is used
    • int: Every sliding_window layers, use global attention, else use local attention.
    • str: A sequence of “L” (local attention) and “G” (global attention) characters that defines the attention pattern. The pattern starts from layer 0 and repeats every sliding_window layers. The final layer always uses global attention regardless of the pattern. For instance, sliding_window_pattern=“LLLG” same as sliding_window=4, which means:
    • Layer 0, 1, 2: local attention,
    • Layer 3: global attention, …(repeated)
  • layer_types (list, optional) — Attention pattern for each layer. Prioritized over sliding_window_pattern.
  • mlp_layer_types (list, optional) — MLP pattern for each layer. Prioritized over first_k_dense_replace.
  • first_k_dense_replace (int, optional, defaults to 1) — Number of dense layers in shallow layers(embed->dense->dense->…->dense->moe->moe…->lm_head). --k dense layers—/
  • moe_intermediate_size (int, optional, defaults to 1024) — Dimension of the MoE representations.
  • num_experts (int, optional, defaults to 64) — Number of routed experts.
  • num_experts_per_tok (int, optional, defaults to 8) — Number of selected experts, None means dense model.
  • num_shared_experts (int, optional, defaults to 1) — Number of shared experts.
  • norm_topk_prob (bool, optional, defaults to True) — Whether to normalize the weights of the routed experts.
  • routed_scaling_factor (float, optional, defaults to 2.5) — Scaling factor or routed experts.
  • n_group (int, optional, defaults to 1) — Number of groups for routed experts.
  • topk_group (int, optional, defaults to 1) — Number of selected groups for each token(for each token, ensuring the selected experts is only within topk_group groups).

This is the configuration class to store the configuration of a [ExaoneMoeModel]. It is used to instantiate a EXAONE MoE model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the K-EXAONE-236B-A23B LGAI-EXAONE/K-EXAONE-236B-A23B

Configuration objects inherit from [PreTrainedConfig] and can be used to control the model outputs. Read the documentation from [PreTrainedConfig] for more information.

Example:

>>> from transformers import ExaoneMoeModel, ExaoneMoeConfig

>>> # Initializing a EXAONE configuration
>>> configuration = ExaoneMoeConfig()

>>> # Initializing a model from configuration
>>> model = ExaoneMoeModel(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config

ExaoneMoeModel

class transformers.ExaoneMoeModel

< >

( config: ExaoneMoeConfig )

Parameters

  • config (ExaoneMoeConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The bare Exaone Moe Model outputting raw hidden-states without any specific head on top.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

< >

( input_ids: torch.LongTensor | None = None attention_mask: torch.Tensor | None = None position_ids: torch.LongTensor | None = None past_key_values: transformers.cache_utils.Cache | None = None inputs_embeds: torch.FloatTensor | None = None use_cache: bool | None = None cache_position: torch.LongTensor | None = None **kwargs: typing_extensions.Unpack[transformers.utils.generic.TransformersKwargs] )

ExaoneMoeForCausalLM

class transformers.ExaoneMoeForCausalLM

< >

( config )

Parameters

  • config (ExaoneMoeForCausalLM) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The Exaone Moe Model for causal language modeling.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

< >

( input_ids: torch.LongTensor | None = None attention_mask: torch.Tensor | None = None position_ids: torch.LongTensor | None = None past_key_values: transformers.cache_utils.Cache | None = None inputs_embeds: torch.FloatTensor | None = None labels: torch.LongTensor | None = None use_cache: bool | None = None cache_position: torch.LongTensor | None = None logits_to_keep: int | torch.Tensor = 0 **kwargs: typing_extensions.Unpack[transformers.utils.generic.TransformersKwargs] ) → transformers.modeling_outputs.CausalLMOutputWithPast or tuple(torch.FloatTensor)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • attention_mask (torch.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

    • 1 for tokens that are not masked,
    • 0 for tokens that are masked.

    What are attention masks?

  • position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.n_positions - 1].

    What are position IDs?

  • past_key_values (~cache_utils.Cache, optional) — Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the past_key_values returned by the model at a previous stage of decoding, when use_cache=True or config.use_cache=True.

    Only Cache instance is allowed as input, see our kv cache guide. If no past_key_values are passed, DynamicCache will be initialized by default.

    The model will output the same cache format that is fed as input.

    If past_key_values are used, the user is expected to input only unprocessed input_ids (those that don’t have their past key value states given to this model) of shape (batch_size, unprocessed_length) instead of all input_ids of shape (batch_size, sequence_length).

  • inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectors than the model’s internal embedding lookup matrix.
  • labels (torch.LongTensor of shape (batch_size, sequence_length), optional) — Labels for computing the masked language modeling loss. Indices should either be in [0, ..., config.vocab_size] or -100 (see input_ids docstring). Tokens with indices set to -100 are ignored (masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size].
  • use_cache (bool, optional) — If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).
  • cache_position (torch.LongTensor of shape (sequence_length), optional) — Indices depicting the position of the input sequence tokens in the sequence. Contrarily to position_ids, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length.
  • logits_to_keep (Union[int, torch.Tensor], optional, defaults to 0) — If an int, compute logits for the last logits_to_keep tokens. If 0, calculate logits for all input_ids (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. If a torch.Tensor, must be 1D corresponding to the indices to keep in the sequence length dimension. This is useful when using packed tensor format (single dimension for batch and sequence length).

Returns

transformers.modeling_outputs.CausalLMOutputWithPast or tuple(torch.FloatTensor)

A transformers.modeling_outputs.CausalLMOutputWithPast or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (ExaoneMoeConfig) and inputs.

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Language modeling loss (for next-token prediction).

  • logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).

  • past_key_values (Cache, optional, returned when use_cache=True is passed or when config.use_cache=True) — It is a Cache instance. For more details, see our kv cache guide.

    Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see past_key_values input) to speed up sequential decoding.

  • hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

    Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.

  • attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

    Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

The ExaoneMoeForCausalLM forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Example:

>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> model = AutoModelForCausalLM.from_pretrained("LGAI-EXAONE/K-EXAONE-236B-A23B")
>>> tokenizer = AutoTokenizer.from_pretrained("LGAI-EXAONE/K-EXAONE-236B-A23B")

>>> prompt = "Explain how wonderful you are"
>>> messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
>>> input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    enable_thinking=False,
)

>>> output = model.generate(**input_ids.to(model.device), max_new_tokens=128)
>>> tokenizer.decode(output[0], skip_special_tokens=False)
"<|system|>\nYou are a helpful assistant.<|endofturn|>\n<|user|>\nExplain how wonderful you are<|endofturn|>\n<|assistant|>\n<think>\n\n</think>\n\nThank you for the kind question! While I can't feel emotions or take pride in the way humans do, I *can* share what makes me uniquely helpful and capable—qualities that many people find wonderful.\n\nHere’s how I can support you:\n\n🌟 **Knowledge at Your Fingertips**  \nI have access to a vast amount of information across countless topics—from science and history to technology and creative writing. Whether you're curious, learning, or solving a problem, I can help explain things clearly and accurately.\n\n💬 **Clear, Helpful Communication**  \nI aim to respond in a way that's easy to understand, whether you need a simple explanation or a detailed analysis. I adapt my tone and depth to match"
Update on GitHub

←EXAONE-4.0 Falcon→