K-EXAONE 2.0

2026-07-31

Date the weights (or, if earlier, the technical report) were first made public, YYYY-MM-DD, as dated by the publisher. A Hugging Face repository creation date is used only when no dated publisher page exists, with a note.

apache-2.0

License of the released weights as declared by the publisher (model card metadata or license file), e.g. apache-2.0, mit, or the name of a custom license.

text

Inferred from the source; see the explanation below.

Input modalities of the released checkpoint as the publisher states them, from: text, image, video, audio, joined with ' + ' in that order ("text", "text + image + video").

750B parameters

Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.

37B parameters

Parameters used per token in the forward pass, as reported by the publisher.

Not disclosed

Active parameters excluding the input embedding and the output head (unembedding). Stated by the publisher, or computed from params.active by subtracting vocab × d_model for each of the embedding and output head that params.active is stated to include (basis inferred, note with the arithmetic). Not disclosed when the scope of params.active is not stated.

748.8B parameters

Calculated: embedding + output head + all experts + other weights, LGAI-EXAONE/K-EXAONE-2.0-750B-A37B@fec0c8d safetensors headers; reproduces stated total (MTP) and active (with embedding and output head)

Parameters of the text backbone counted from the released checkpoint's safetensors headers (embedding, output head, all experts and other weights; excluding MTP modules, vision/audio towers and n-gram lookup tables). Shown only when the counts reproduce the publisher's stated total and active parameters within 5% under some scope.

1.887B parameters

Calculated: input embedding + output head, LGAI-EXAONE/K-EXAONE-2.0-750B-A37B@fec0c8d safetensors headers

Input embedding plus output head (once if tied), counted from the checkpoint.

746.9B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, LGAI-EXAONE/K-EXAONE-2.0-750B-A37B@fec0c8d safetensors headers; reproduces stated total (MTP) and active (with embedding and output head)

Language backbone weights excluding input embeddings, output head, lookup/Engram tables, MTP modules and vision/audio towers. Equals params.counted_total minus params.counted_embedding. Same validation as params.counted_total.

9.816B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, LGAI-EXAONE/K-EXAONE-2.0-750B-A37B@fec0c8d safetensors headers

Largest group of language-backbone weights sharing a full layers.<index> tensor path, including every routed expert. Excludes lookup/Engram tables, embeddings, output head, MTP and other modalities. Hybrid checkpoints may store mixers and FFNs as separate layers. All-gather volume also depends on wrapping, sharding and dtype.

35.45B parameters

Calculated: non-expert weights + routed experts × 8/256, excluding embedding and output head, LGAI-EXAONE/K-EXAONE-2.0-750B-A37B@fec0c8d safetensors headers; reproduces stated total (MTP) and active (with embedding and output head)

Active parameters per token excluding the input embedding and output head, counted from the checkpoint as all non-expert weights plus routed-expert weights × top_k / experts. Same validation as params.counted_total.

Main-block count; the MTP block adds 0.52B and the separately trained DSpark drafter 2.53B

What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).

78

Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.

6,144

Residual stream width.

153,600

Embedding table rows (vocab_size in config, may include padding).

No

Input embedding and output head share weights.

inherited unchanged from K-EXAONE

Tokenizer type and lineage as described by the publisher (e.g. byte-level BPE shared with an earlier model).

78.7692

Calculated: d_model / layers

Width-to-depth ratio, d_model divided by the number of blocks.

GQA, SWA

Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).

1 global + 1 SWA (4096 window) + 19 × (3 SWA : 1 global)

Inferred from the source; see the explanation below.

Token-mixer order, summarized as a pattern or layer positions. The fact's optional sequence lists each mixer in order; a ratio alone does not establish the order. Auxiliary prediction modules and standalone FFNs are excluded.

20

Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.

2.9

Calculated: attn.local_layers / attn.global_layers

attn.local_layers divided by attn.global_layers, the sliding-window to global interleave ratio (5 for '5 SWA : 1 global').

58

Layers with a fixed attention window.

64

Query heads in the global-attention layers.

8

Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.

8,192

Calculated: heads × head_dim

attn.heads × attn.head_dim, the query projection width of global layers.

8

Calculated: heads / kv_heads

attn.heads divided by attn.kv_heads in global layers.

128

Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).

128 tokens

Sliding-window size of the local layers.

1

Sliding-window layers (counted in attn.local_layers) that use attn.long_window instead of attn.window.

4.096K tokens

Window of the attn.long_window_layers.

8

KV heads in sliding-window layers (stated even when equal to attn.kv_heads).

64

Query heads in sliding-window layers (stated even when equal to attn.heads).

128

Query/key head dimension in sliding-window layers. Value head dimensions may differ; see attn.local_v_head_dim.

RoPE (SWA), NoPE (global)

Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.

1,000,000

RoPE base frequency of the sliding-window layers (stated even when equal to attn.rope_theta).

Yes

Queries and keys are normalized per head before the dot product.

No

Query/key/value projections have bias terms.

SwiGLU

FFN activation (SwiGLU, GeGLU, ...).

18,432

Intermediate width of the dense FFN layers that are actually executed.

Clamped SwiGLU, limit 7.0 on the last 16 MoE layers: gate clamped above, up clamped both sides

Clamping of GLU gate / up activations (e.g. SwiGLU limit 7), with the layers it applies to.

No

FFN projections have bias terms.

3

Calculated: ffn.dim / d_model

ffn.dim divided by shape.d_model.

256

Routed experts per MoE layer.

8

Routed experts selected per token.

1

Always-on experts per MoE layer.

2,048

Intermediate width of one routed expert.

2,048

Total intermediate width of the shared expert(s).

2

Number of initial blocks that use a dense FFN instead of MoE.

sigmoid

Gating function over expert logits (softmax, sigmoid).

sequence-level load balancing plus DeepSeek-style per-expert score bias; dropless

Load-balancing method, with coefficients where disclosed.

2.5

Multiplier applied to the combined routed-expert output (routed_scaling_factor).

Yes

Selected expert weights are renormalized to sum to 1.

0.333333

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

32

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.0493333

Calculated: params.active / params.total

params.active divided by params.total.

1

Additional future tokens predicted during pretraining (0 when the model was trained without MTP).

1

MTP modules with distinct weights. 1 with mtp.depth > 1 means one module is applied recursively.

RMSNorm on embedding and on previous hidden state, concatenated and projected (fc 12288 → 6144); one SWA layer (window 128, 64/8 heads, QK norm) with dense SwiGLU FFN; final RMSNorm; shared LM head

Contents of one MTP module (mixer and FFN type, norms, how the previous hidden state and next-token embedding are combined).

dense

FFN in the MTP module: 'dense' or 'MoE', with its width when it differs from the backbone.

Yes

MTP modules use the main model's input embedding and output head.

520M parameters

Parameters of all MTP modules as reported by the publisher.

Not disclosed

Weight of the MTP loss, including its schedule.

Yes

The released checkpoint contains the MTP module weights.

RMSNorm

Normalization layer type in the blocks.

pre

Pre-norm, post-norm, or sandwich (pre and post).

1e-05

Epsilon of the block normalization layers.

8.192K tokens

Sequence length of the main pretraining stage.

262.1K tokens

Longest sequence length trained on after context extension.

262.1K tokens

Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.

mid-training in two stages: 8K → 64K (400B tokens) → 256K (400B tokens)

How context was extended (stages, RoPE base change, YaRN factor).

8T tokens

Tokens seen in pretraining including annealing and context extension, as reported.

Not disclosed

Optimizer used for the bulk of parameters.

Not disclosed

Peak learning rate of the main pretraining stage.

Not disclosed

Learning rate at the end of the decay.

Not disclosed

Learning-rate schedule shape with phase boundaries.

Not disclosed

Warmup length in steps or tokens.

Not disclosed

Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.

Not disclosed

Decoupled weight decay coefficient.

Not disclosed

(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.

Not disclosed

Global gradient-norm clipping threshold.

Not disclosed

Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.

upcycled from K-EXAONE: depth 48 → 78 layers (12 → 19 LLLG blocks, repeating blocks from the middle of the stack); width 128 → 256 experts (each expert duplicated with its router row, plus norm-preserving random rotation noise)

Weight initialization (std or scheme) as stated by the publisher; config initializer_range alone is not evidence.

Clamped SwiGLU (limit 7.0) on the last 16 layers

Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.

K-EXAONE mixture plus Active Reading synthetic data, institution-sourced Korean culture and history corpora, and FineWeb2 for four added languages

Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).

2025 2Q

Pretraining data cutoff date.

• healing

  Source: paper · p.4-7 §2.2, §3, §4

  To maintain training stability following this expansion, we first perform a healing stage using a portion of the later-stage K-EXAONE training data.

  Starts from upcycled K-EXAONE weights (train.init); the report says depth up-scaling is applied during the early stage of pre-training (p.5 §3). Healing-stage token count is not disclosed; mid-training extends the context from 8K (p.7 §4).

• continual pre-training

  Tokens

  8000000000000

  Source: paper · p.4-7 §2.2, §3, §4

  After healing, the model is further trained on an additional 8T tokens using the K-EXAONE 2.0 pre-training data mixture.

  Starts from upcycled K-EXAONE weights (train.init); the report says depth up-scaling is applied during the early stage of pre-training (p.5 §3). Healing-stage token count is not disclosed; mid-training extends the context from 8K (p.7 §4).

• mid-stage 1

  Tokens

  400000000000

  Seq len

  65536

  Source: paper · p.4-7 §2.2, §3, §4

  In the first stage (Mid-Stage 1), the model is trained on 400B tokens with a 64K context window

  Starts from upcycled K-EXAONE weights (train.init); the report says depth up-scaling is applied during the early stage of pre-training (p.5 §3). Healing-stage token count is not disclosed; mid-training extends the context from 8K (p.7 §4).

• mid-stage 2

  Tokens

  400000000000

  Seq len

  262144

  Source: paper · p.4-7 §2.2, §3, §4

  followed by training on an additional 400B tokens with context windows exceeding 64K tokens in the second stage (Mid-Stage 2).

  Context windows above 64K, up to 256K.

Ordered pretraining stages. Each item: name, tokens, seq_len, lr, batch_tokens, batch_sequences, steps, note, quote, and source when it differs from the fact's source.

small-scale K-EXAONE model (size not disclosed) for mid-training stage ablations

Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.

10.6667

Calculated: train.tokens / params.total

train.tokens divided by params.total.

225.649

Calculated: train.tokens / params.counted_active_non_embedding

train.tokens divided by params.counted_active_non_embedding, or by the stated params.active_non_embedding when there is no validated checkpoint count.

82,276 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 20 × (2048 × 16/8) B/entry; SWA KV: 57 × (2048 × 16/8) B/entry; Long-window KV: 1 × (2048 × 16/8) B/entry

Retained KV cache divided by context length, including bounded SWA caches and indexer keys. Exports use uniform BF16 at 131,072 tokens. The website can select uniform payload precision (no quantization overhead) or sourced component formats (including scales and documented padding). This is cache payload, not peak prefill memory or persistent prefix storage.

81,920 bytes

Calculated: 81920 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 20 × (2048 × 16/8) B/entry; SWA KV: 57 × (2048 × 16/8) B/entry; Long-window KV: 1 × (2048 × 16/8) B/entry

Long-context global KV bytes per token, including indexer keys; excludes bounded SWA caches. Compression rates and cross-layer cache reuse are included. Exports use uniform BF16; the website's storage selector also applies here. This is the field to compare with global-KV figures in reports, using the matching storage recipe.

2,224 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 20 × (2048 × 16/8) B/entry; SWA KV: 57 × (2048 × 16/8) B/entry; Long-window KV: 1 × (2048 × 16/8) B/entry

kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).

291,461,618 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 20 × (2048 × 16/8) B/entry; SWA KV: 57 × (2048 × 16/8) B/entry; Long-window KV: 1 × (2048 × 16/8) B/entry

kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).

10,784,079,872 bytes

Calculated: 81920 B × 131072 (Global KV) + 233472 B × min(131072, 128) (SWA KV) + 4096 B × min(131072, 4096) (Long-window KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 20 × (2048 × 16/8) B/entry; SWA KV: 57 × (2048 × 16/8) B/entry; Long-window KV: 1 × (2048 × 16/8) B/entry

Retained KV cache for one sequence at the selected context, including SWA windows and indexer keys. Exports use uniform BF16 at 131,072 tokens. Excludes recurrent state, compressor tail state, draft/MTP caches, temporary prefill tensors, allocator metadata and persistent prefix snapshots. Does not estimate peak prefill or total serving memory.

20

Artificial Analysis Intelligence Index v4.3.2 of the post-trained model, from a dated Wayback snapshot of its Artificial Analysis page. The note names the variant (reasoning setting).