Laguna-XS.2

2026-04-28

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").

33.4B parameters

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

3B parameters

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

2.59B parameters

Inferred from the source; see the explanation below.

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.

33.44B parameters

Calculated: embedding + output head + all experts + other weights, poolside/Laguna-XS.2@a397bde safetensors headers; reproduces stated total (backbone) 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.

411M parameters

Calculated: input embedding + output head, poolside/Laguna-XS.2@a397bde safetensors headers

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

33.03B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, poolside/Laguna-XS.2@a397bde safetensors headers; reproduces stated total (backbone) 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.

846.9M parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, poolside/Laguna-XS.2@a397bde 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.

2.606B parameters

Calculated: non-expert weights + routed experts × 8/256, excluding embedding and output head, poolside/Laguna-XS.2@a397bde safetensors headers; reproduces stated total (backbone) 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.

Active count includes embeddings.

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

40

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

2,048

Residual stream width.

100,352

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

No

Input embedding and output head share weights.

BPE trained on internal data; shared with M.1

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

51.2

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 : 3 SWA

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.

10

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

3

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').

30

Layers with a fixed attention window.

48

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.

6,144

Calculated: heads × head_dim

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

6

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

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

512 tokens

Sliding-window size of the local 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.

partial RoPE (global, first 50%), RoPE (SWA)

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

500,000

RoPE base frequency of the global layers at release.

10,000

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

0.5

Fraction of each head's QK dimensions that are rotated.

Yes

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

softplus, per head, on attention output before o_proj

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

No

Query/key/value projections have bias terms.

SwiGLU

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

8,192

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

No

FFN projections have bias terms.

4

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.

512

Intermediate width of one routed expert.

512

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-batch auxiliary loss (Qiu et al.) over non-padding tokens; padding tokens skip routing

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.

0.25

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.

0.0898204

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).

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.

YaRN on global layers only; 4K → 32K → 128K (100B tokens each); 256K by doubling the global RoPE scale without training

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

>30T

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

Muon + AdamW (Moonlight variant; AdamW for non-matrix parameters)

Optimizer used for the bulk of parameters.

0.0005

Peak learning rate of the main pretraining stage.

2.5e-05

Learning rate at the end of the decay.

WSD: linear warmup, stable at peak, 1−√ decay over the last 30% of steps to 5% of peak

Learning-rate schedule shape with phase boundaries.

Not disclosed

Warmup length in steps or tokens.

24M tokens

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

24M tokens

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.

Not disclosed

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

Not disclosed

Global gradient-norm clipping threshold.

BF16 mixed precision, FP32 master weights; selective FP32 in RMSNorm, RoPE and the LM-head input-gradient all-reduce

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

FP32 LM-head input-gradient all-reduce; no z-loss; padding tokens skip routing and load balancing

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

document masking; position ids reset at document boundaries

Whether attention crosses document boundaries within a packed sequence.

AutoMixer mix: raw code 30.6%, web 25.2%, synthetic/code-text 25.4%, math 9.0%, knowledge 6.6%, instruction-like 1.4%, academic papers 1.1%, books 0.7%; sampled from a ~27T unique-token pool

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

2,048 H200

Accelerators used for pretraining, count and duration where stated.

2,048 H200

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").

EP8

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").

EMA over the 10 most recent checkpoints at the end of the 128K stage

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

• main

  Seq len

  4096

  Lr

  0.0005

  Batch tokens

  24000000

  Source: paper · p.6 §3.1.1

  The pre-training stage uses a context length of 4K tokens

  WSD; tokens not disclosed

• extend-32K

  Tokens

  100000000000

  Seq len

  32768

  Lr

  2.5e-05

  Batch tokens

  24000000

  Source: paper · p.6 §3.1.1

  split it into two equal-token sub-stages of 100B tokens each. The first sub-stage extends the context to 32K tokens

  cosine 2.5e-5 → 5e-6, no re-warmup; YaRN on global layers

• extend-128K

  Tokens

  100000000000

  Seq len

  131072

  Lr

  2.5e-05

  Batch tokens

  24000000

  Source: paper · p.6 §3.1.1

  the second extends the context to 128K tokens

  cosine 2.5e-5 → 5e-6; EMA of the last 10 checkpoints is the final base

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.

peak LR from WSD scaling law lr⋆(N, D) = 10^4.488 · N^−0.4639 · D^−0.2661 (N active params, D tokens incl. cooldown), scaled by √(B/B0) from B0 = 8M; predicts ~5.5e-4 at N = 3.0B, B = 24M; 5e-4 used

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.

LR-law sweep: 4 MoE sizes (2B–16B total / 0.3B–2.2B active) × 6 LRs × 5 token budgets (30B–480B) at 8M batch; architecture ablations on a 16.6B-total / 2.3B-active proxy (22 layers, d_model 2048, 128 experts top-8, 335B tokens); data mixture from ~60 × ~0.5B MoE proxies on ~60B tokens

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

41,440 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry; SWA KV: 30 × (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.

40,960 bytes

Calculated: 40960 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry; SWA KV: 30 × (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.

13,813 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry; SWA KV: 30 × (2048 × 16/8) B/entry

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

1,810,541,227 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry; SWA KV: 30 × (2048 × 16/8) B/entry

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

5,431,623,680 bytes

Calculated: 40960 B × 131072 (Global KV) + 122880 B × min(131072, 512) (SWA KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry; SWA KV: 30 × (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.