A.X K2

2026-07-29

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

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").

688B parameters

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

33B 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.

691.7B parameters

Calculated: embedding + output head + all experts + other weights, skt/A.X-K2@2287ca4 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

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.

2.349B parameters

Calculated: input embedding + output head, skt/A.X-K2@2287ca4 safetensors headers

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

689.3B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, skt/A.X-K2@2287ca4 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

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.

34.02B parameters

Calculated: non-expert weights + routed experts × 8/256, excluding embedding and output head, skt/A.X-K2@2287ca4 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

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.

61

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

7,168

Residual stream width.

163,840

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

No

Input embedding and output head share weights.

byte-level BPE, 163,840 tokens, inherited unchanged from A.X K1

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

117.508

Calculated: d_model / layers

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

MLA, DSA (SGA)

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

61 MLA

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.

61

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

64

Query heads in the global-attention layers.

12,288

Calculated: heads × head_dim

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

192

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

128

Value dimension per head in global layers (MLA v_head_dim; stated for MLA even when equal to the QK head dim).

512

Width of the compressed KV latent that MLA caches per token per layer.

1,536

Width of the low-rank query projection in MLA; 0 when queries use a full-rank projection.

64

Width of the shared positional key that MLA caches next to the latent.

partial RoPE (64-dim decoupled RoPE key shared across heads, 128 NoPE dims per head)

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

1,000,000

RoPE base frequency of the global layers at release.

No

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

head-specific sigmoid gate on the attention output before o_proj, projected jointly with the query

Gate applied to the attention output before the output projection, with its placement.

DSA-style indexer, top-2048 tokens per query; 64 indexer heads of dim 128, ReLU-scored

Token-selection mechanism inside global layers, e.g. DSA lightning indexer with top-k.

No

Query/key/value projections have bias terms.

128

Width of the sparse-attention indexer key cached per token in each global layer (DSA lightning indexer), counted in the KV cache.

SwiGLU

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

18,432

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

No

FFN projections have bias terms.

2.57143

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).

1

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

sigmoid

Gating function over expert logits (softmax, sigmoid).

global auxiliary loss (4e-3 general and reasoning stages, 1e-3 long-context, 5e-4 SGA adaptation) plus a learnable expert bias (update rate 1e-3, lowered to 1e-4)

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.

256 experts in 8 groups, top-4 groups per token

Expert groups and groups selected per token, or node-limited routing, when used.

0.285714

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.

0.0479651

Calculated: params.active / params.total

params.active divided by params.total.

Not disclosed

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

No

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-06

Epsilon of the block normalization layers.

4.096K tokens

Sequence length of the main pretraining stage.

131.1K 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.

ABF: RoPE base 1e4 through stages 1-2, raised to 1e6 for the 32K and 128K stages; 256K at inference by raising the YaRN factor

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

8.2T tokens

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

AdamW (eps 1e-8)

Optimizer used for the bulk of parameters.

0.00022

Peak learning rate of the main pretraining stage.

2.2e-05

Learning rate at the end of the decay.

WSD: warmup over ~25B tokens, constant 2.2e-4 through stage 1, cosine to 7e-5 across stage 2; 32K stage re-warms to 7e-5 and cosine-decays to 2.2e-5; 128K stage warms to and holds 2.2e-5

Learning-rate schedule shape with phase boundaries.

~6.1M sequences (~25B tokens)

Warmup length in steps or tokens.

16,384 × 4,096 (67.1M tokens); ramped from 2,048 sequences in steps of 2,048 over the first 72M sequences (~295B tokens)

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

67.11M tokens

Inferred from the source; see the explanation below.

Global batch size in tokens during the main pretraining stage, after any ramp (sequences × sequence length when stated in sequences, marked inferred).

Not disclosed

Decoupled weight decay coefficient.

(0.9, 0.95)

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

Not disclosed

Global gradient-norm clipping threshold.

MXFP8 (E4M3) forward and backward, one scale per 32 consecutive elements; FP32 master weights and gradients, BF16 optimizer moments; FP8 parameter all-gather

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

Gated Norm (sigmoid gate after the block RMSNorms), replacing the dual-normalization scheme of A.X K1

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

web, code, academic papers, STEM, books, Q&A and synthetic; stage-1 designed mix English 72.7%, Korean 15.4%, code 8.3%

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

512 NVIDIA B200 GPUs (64 nodes of 8) for ~70 days

Accelerators used for pretraining, count and duration where stated.

512 B200

Accelerator count and type for the main pretraining run, as the count with thousands separators and the chip's short name ("2,048 H800", "2,048 GH200"). A stated range or bound stays as stated ("2,048–4,096 GH200").

PP8 · VPP2 · EP8 · DP (distributed optimizer)

Parallelism layout of the main pretraining run as stated, as degree-annotated terms in the order TP, PP, VPP, EP, CP, DP ("PP16 · EP64 · ZeRO-1 DP"); a term without a stated degree has none ("FSDP").

weight-space average of the last 6 checkpoints, ~1.6B tokens apart, with 1-sqrt weighting

EMA or checkpoint merging applied to produce the released base weights.

• stage 1 - general knowledge

  Tokens

  6400000000000

  Seq len

  4096

  Source: paper · p.6-7 §3.3, §3.4

  A.X K2 is trained on a corpus of 6.4 trillion tokens at a sequence length of 4,096

  Stage 3C adds the sparse indexer: an indexer-only sparse warmup with the backbone frozen, then joint training with the indexer loss down-weighted.

• stage 2 - high-quality reasoning

  Tokens

  1400000000000

  Seq len

  4096

  Source: paper · p.6-7 §3.3, §3.4

  Using a curated corpus of 1.4 trillion tokens at a sequence length of 4,096, this stage is strictly filtered for high-complexity web data

  Stage 3C adds the sparse indexer: an indexer-only sparse warmup with the backbone frozen, then joint training with the indexer loss down-weighted.

• stage 3A - 32K extension

  Tokens

  ~0.22T

  Seq len

  32768

  Source: paper · p.6-7 §3.3, §3.4

  we natively extend the context length to 32K (Stage 3A, approximately 0.22T tokens)

  Stage 3C adds the sparse indexer: an indexer-only sparse warmup with the backbone frozen, then joint training with the indexer loss down-weighted.

• stage 3B/3C - 128K extension and SGA adaptation

  Tokens

  ~0.14T

  Seq len

  131072

  Source: paper · p.6-7 §3.3, §3.4

  Stages 3B and 3C together consume approximately 0.14T tokens.

  Stage 3C adds the sparse indexer: an indexer-only sparse warmup with the backbone frozen, then joint training with the indexer loss down-weighted.

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.

LR and batch magnitudes from the MoE scaling laws of Tian et al. (2025)

How peak LR, batch size or other hyperparameters were chosen for the released run (fitted scaling law with its formula, μP transfer from a proxy width, sweep), as stated.

20B-A3B dry run at sequence length 4096 (Gated Norm ablation)

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

11.9186

Calculated: train.tokens / params.total

train.tokens divided by params.total.

241.011

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.

85,888 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (704 × 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.

85,888 bytes

Calculated: 85888 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (704 × 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.

2,603 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (704 × 16/8) B/entry

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

341,136,725 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (704 × 16/8) B/entry

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

11,257,511,936 bytes

Calculated: 85888 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (704 × 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.

21

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).