Laguna-M.1-base
Poolside · base checkpoint
Open in the comparison table Hugging Face ↗ Paper ↗
KV calculations use one sequence at 131,072 tokens and 16-bit cache elements. They include bounded sliding-window caches. This is retained cache payload, not peak prefill memory. The comparison table also offers sourced mixed-precision layouts, including quantization scales.
Release
- Released
2026-05-25
Source: paper · p.1 title block
RESEARCH@POOLSIDE.AI MAY 25, 2026
Technical report date; M.1 was API-only from 2026-04-28 and its weights were published after the report.
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.
- License
apache-2.0
Source: model-card · README.md; YAML front matter
license: 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.
- Input modalities
text
Source: model-card · Model overview
Modality: text-to-text (completion)
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").
Parameters
- Total params
225.8B parameters
Source: paper · p.5 §3.1 Architecture & Training
LAGUNA M.1 has 225.8B total parameters
Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.
- Active params
23.4B parameters
Source: paper · p.5 §3.1 Architecture & Training
out of which 23.4B (including embeddings) are activated per token
Parameters used per token in the forward pass, as reported by the publisher.
- Active non-embedding (stated)
22.6B parameters
Inferred from the source; see the explanation below.
Source: paper · p.5 §3.1 Architecture & Training
out of which 23.4B (including embeddings) are activated per token
23.4B − 2 × 100,352 × 4,096 (input embedding and untied LM head) = 22.58B, rounded to the 0.1B precision of the stated count.
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.
- Counted total params
225.8B parameters
Calculated: embedding + output head + all experts + other weights, poolside/Laguna-M.1-base@628037a 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.
- Counted embedding + head
822.1M parameters
Calculated: input embedding + output head, poolside/Laguna-M.1-base@628037a safetensors headers
Input embedding plus output head (once if tied), counted from the checkpoint.
- Non-emb. params (counted)
225B parameters
Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, poolside/Laguna-M.1-base@628037a 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.
- Largest layer (counted)
3.344B parameters
Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, poolside/Laguna-M.1-base@628037a 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.
- Active non-emb. (counted)
22.64B parameters
Calculated: non-expert weights + routed experts × 16/256, excluding embedding and output head, poolside/Laguna-M.1-base@628037a 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.
- Counting scope
Active count includes embeddings.
Source: paper · p.5 §3.1 Architecture & Training
out of which 23.4B (including embeddings) are activated per token
What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).
Shape
- Layers
70
Source: config · /num_hidden_layers
Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.
- d_model
4,096
Source: config · /hidden_size
Residual stream width.
- Vocab
100,352
Source: config · /vocab_size
Embedding table rows (vocab_size in config, may include padding).
- Tied embeddings
No
Source: config · /tie_word_embeddings
Input embedding and output head share weights.
- Tokenizer
BPE trained on internal data; shared with XS.2
Source: paper · p.5 §3.1
Both models share the same tokenizer with a vocabulary size of 100,352 tokens, trained with Byte-Pair Encoding [75] on our internal datasets.
Tokenizer type and lineage as described by the publisher (e.g. byte-level BPE shared with an earlier model).
- d_model / layers
58.5143
Calculated: d_model / layers
Width-to-depth ratio, d_model divided by the number of blocks.
Attention
- Token mixers
GQA
Source: model-card · README.md; Highlights
Laguna M.1 uses global attention across all layers with 64 Q-heads, 8 KV-heads
Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).
- Layer layout
all global
Source: model-card · README.md; Highlights
Laguna M.1 uses global attention across all layers
Config sliding_window 0 means SWA is disabled.
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.
- Global-attention layers
70
Source: model-card · README.md; Model overview
Layers: 70 layers with global attention
Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.
- Query heads
64
Source: config · /num_attention_heads
Query heads in the global-attention layers.
- KV heads
8
Source: config · /num_key_value_heads
Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.
- Heads × head dim
8,192
Calculated: heads × head_dim
attn.heads × attn.head_dim, the query projection width of global layers.
- Q heads / KV head
8
Calculated: heads / kv_heads
attn.heads divided by attn.kv_heads in global layers.
- QK head dim
128
Source: config · /head_dim
Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).
- V head dim
128
Source: config · /head_dim
Value dimension per head in global layers (MLA v_head_dim; stated for MLA even when equal to the QK head dim).
- Positional encoding
RoPE
Source: model-card · README.md; Model overview
Positional encoding: RoPE with YaRN
Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.
- RoPE base
500,000
Source: config · /rope_parameters/full_attention/rope_theta
RoPE base frequency of the global layers at release.
- Rotary fraction
1
Source: config · /rope_parameters/full_attention/partial_rotary_factor
Fraction of each head's QK dimensions that are rotated.
- QK norm
Yes
Source: code · Laguna-M.1 modeling_laguna.py; LagunaAttention.__init__
self.q_norm = LagunaRMSNorm(config.head_dim, eps=config.rms_norm_eps)
Queries and keys are normalized per head before the dot product.
- Output gate
softplus, per element, on attention output before o_proj
Source: model-card · README.md; Model overview
Attention: 64 Q-heads, 8 KV-heads, head dimension 128, with softplus attention output gating
Per element: config /gating = "per-element" (one gate per head channel); XS.2 moved to per-head.
Gate applied to the attention output before the output projection, with its placement.
- QKV bias
No
Source: code · Laguna-M.1 modeling_laguna.py; LagunaAttention.__init__
self.q_proj = nn.Linear(config.hidden_size, self.num_heads * config.head_dim, bias=False)
Query/key/value projections have bias terms.
Feed-forward
- Activation
SwiGLU
Source: model-card · README.md; Model overview
Dense layers: first 3 layers are dense SwiGLU; remaining 67 layers are sparse MoE
FFN activation (SwiGLU, GeGLU, ...).
- Dense FFN dim
16,384
Source: config · /intermediate_size
Intermediate width of the dense FFN layers that are actually executed.
- FFN bias
No
Source: code · Laguna-M.1 modeling_laguna.py; LagunaMLP.__init__
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
FFN projections have bias terms.
- Dense FFN / d_model
4
Calculated: ffn.dim / d_model
ffn.dim divided by shape.d_model.
Mixture of experts
- Routed experts
256
Source: config · /num_experts
Routed experts per MoE layer.
- Experts per token
16
Source: config · /num_experts_per_tok
Routed experts selected per token.
- Expert dim
1,024
Source: config · /moe_intermediate_size
Intermediate width of one routed expert.
- Leading dense layers
3
Source: config · /mlp_only_layers | len
Layers 0-2; decoder_sparse_step 1 makes every other layer MoE (LagunaDecoderLayer.__init__).
Number of initial blocks that use a dense FFN instead of MoE.
- Router score
sigmoid
Source: code · Laguna-M.1 modeling_laguna.py; LagunaTopKRouter
Laguna MoE router using sigmoid scoring (not softmax).
Gating function over expert logits (softmax, sigmoid).
- Load balancing
auxiliary-loss-free
Source: model-card · README.md; Highlights
Laguna M.1 uses 67 sparse MoE layers with 256 experts, top-k=16 routing and auxiliary-loss-free load balancing
Report §3.1.3 says M.1 padding tokens were "included in the load-balancing loss"; coefficient not disclosed.
Load-balancing method, with coefficients where disclosed.
- Routed scaling factor
1
Source: config · /moe_routed_scaling_factor
Routed-expert modulation (2.5) was introduced in XS.2 (report §3.1).
Multiplier applied to the combined routed-expert output (routed_scaling_factor).
- Renormalize top-k
Yes
Source: config · /norm_topk_prob
Selected expert weights are renormalized to sum to 1.
- Expert dim / d_model
0.25
Calculated: moe.expert_dim / d_model
moe.expert_dim divided by shape.d_model.
- Sparsity (experts / top-k)
16
Calculated: experts / top_k
Routed experts divided by experts selected per token.
- Active / total
0.103632
Calculated: params.active / params.total
params.active divided by params.total.
Multi-token prediction
- MTP depth
Not disclosed
Additional future tokens predicted during pretraining (0 when the model was trained without MTP).
Normalization and stability
- Norm
RMSNorm
Source: paper · p.5 §3.1
with RMSNorm [103] as the normalization layer
Normalization layer type in the blocks.
- Norm placement
pre
Source: paper · p.5 §3.1
LAGUNA XS.2 and M.1 are both Mixture-of-Experts (MoE) [77] models using a pre-norm Transformer architecture
Pre-norm, post-norm, or sandwich (pre and post).
- Norm epsilon
1e-06
Source: config · /rms_norm_eps
Epsilon of the block normalization layers.
Context length
- Pretraining context
Not disclosed
Sequence length of the main pretraining stage.
- Trained context
Not disclosed
Longest sequence length trained on after context extension.
- Advertised context
262.1K tokens
Source: model-card · README.md; Model overview
Context window: 262,144 tokens
Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.
- Extension method
YaRN factor 64 over original 4,096 (release config)
Source: config · config.json; rope_parameters.full_attention
"rope_type": "yarn", "factor": 64.0, "original_max_position_embeddings": 4096,
M.1's long-context training stages are not disclosed; the report's 32K/128K recipe is stated for XS.2. Same dict: beta_fast 64, beta_slow 1, attention_factor 1.416; max_position_embeddings 262,144.
How context was extended (stages, RoPE base change, YaRN factor).
Pretraining recipe
- Pretraining tokens
>30T
Source: paper · p.9 §3.2
In total, both models were trained on more than 30T tokens.
Tokens seen in pretraining including annealing and context extension, as reported.
- Optimizer
Muon + AdamW (AdamW for non-matrix parameters)
Source: paper · p.8 §3.1.3 Expert Collapse
To align the effective weight decay between matrix parameters (updated by Muon)
Same sentence: non-matrix parameters updated by AdamW. Muon in M.1 pre-training: §3.1.2 Distributed Muon reports its overhead "during LAGUNA M.1 pre-training".
Optimizer used for the bulk of parameters.
- Peak LR
Not disclosed
Peak learning rate of the main pretraining stage.
- Final LR
Not disclosed
Learning rate at the end of the decay.
- LR schedule
cosine
Inferred from the source; see the explanation below.
Source: paper · p.5 §3.1 From Laguna M.1 to Laguna XS.2
Second, we adopted a Warmup-Stable-Decay (WSD) learning rate schedule [34] instead of cosine.
Listed as an M.1 → XS.2 change, implying M.1 used cosine. Phase boundaries not disclosed.
Learning-rate schedule shape with phase boundaries.
- Warmup
Not disclosed
Warmup length in steps or tokens.
- Batch size
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.
- Weight decay
Not disclosed
Decoupled weight decay coefficient.
- Adam betas
Not disclosed
(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.
- Grad clip
Not disclosed
Global gradient-norm clipping threshold.
- Training precision
Not disclosed
Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.
- Stability measures
FP32 LM-head input-gradient all-reduce to fix positive logit drift; no z-loss
Source: paper · p.9 §3.1.3
We therefore enforce the LM head input-gradient all-reduce in FP32 while keeping the layer tensor-parallel, resolving this numerical instability.
Introduced after M.1 logit drift; the report does not say at what point in pretraining. Same paragraph: "no z-loss".
Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.
- Document masking
document masking; position ids reset at document boundaries
Source: paper · p.6 §3.1.2 Model Factory component: Titan
resetting position indices at document boundaries to avoid RoPE-related rounding under document masking in attention
Whether attention crosses document boundaries within a packed sequence.
- Data mix
sampled from a ~27T unique-token pool (web, code, educational, academic, synthetic); manually designed mixture with a high-precision web pipeline
Source: paper · p.9 §3.2
We trained both LAGUNA M.1 and LAGUNA XS.2 from scratch on a pre-training corpus sampled from a pool of ∼27T unique tokens.
Same section: for M.1 "we used a high-precision pipeline with a manually designed mixture"; the Table 4 AutoMixer mix is XS.2's.
Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).
- Hardware
6,144 H200
Source: paper · p.6 §3.1.1
LAGUNA M.1 and LAGUNA XS.2 were pre-trained on 6,144 and 2,048 NVIDIA H200 GPUs, respectively.
Accelerators used for pretraining, count and duration where stated.
- Accelerators
6,144 H200
Source: paper · p.6 §3.1.1
LAGUNA M.1 and LAGUNA XS.2 were pre-trained on 6,144 and 2,048 NVIDIA H200 GPUs, respectively.
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").
- Parallelism
EP8
Source: paper · p.7 §3.1.2
For MoE layers, we use Expert Parallelism, sharding expert weights across 8 H200 GPUs interconnected via NVLink.
Titan's mesh is (PP, DDP, FSDP, EGP, ETP) for MoE layers with ETP = 1 for M.1 and XS.2; other degrees not stated.
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").
KV cache
- KV bytes / token
286,720 bytes
Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 70 × (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.
- Global KV bytes / token
286,720 bytes
Calculated: 286720 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 70 × (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.
- KV bytes / token / B active
12,253 bytes
Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 70 × (2048 × 16/8) B/entry
kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).
- KV at context / B active
1,606,024,096 bytes
Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 70 × (2048 × 16/8) B/entry
kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).
- KV at context
37,580,963,840 bytes
Calculated: 286720 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 70 × (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.