Trinity-Large-Base

2026-01-27

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.

other (OpenMDW-1.1)

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").

398B parameters

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

13B 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.

398.6B parameters

Calculated: embedding + output head + all experts + other weights, arcee-ai/Trinity-Large-Base@b2a665b safetensors headers; reproduces stated total (backbone) and active (with 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.

1.23B parameters

Calculated: input embedding + output head, arcee-ai/Trinity-Large-Base@b2a665b safetensors headers

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

397.4B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, arcee-ai/Trinity-Large-Base@b2a665b safetensors headers; reproduces stated total (backbone) and active (with 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.

7.34B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, arcee-ai/Trinity-Large-Base@b2a665b 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.

12.14B parameters

Calculated: non-expert weights + routed experts × 4/256, excluding embedding and output head, arcee-ai/Trinity-Large-Base@b2a665b safetensors headers; reproduces stated total (backbone) and active (with 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.

Card gives ~398B total, report 400B; embedding scope not stated.

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

60

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

3,072

Residual stream width.

200,192

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

No

Input embedding and output head share weights.

byte-level BPE, 200K vocab trained on ~10B tokens; DeepSeek-V3-style pretokenizer with right-aligned 3-digit number chunks and extended script isolation

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).

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.

15

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').

45

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).

4.096K tokens

Sliding-window size of the local layers.

8

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

48

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.

10,000

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.

elementwise sigmoid gate σ(W_G x) on the SDPA output before o_proj, all layers

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, ...).

12,288

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.

4

Routed experts selected per token.

1

Always-on experts per MoE layer.

3,072

Intermediate width of one routed expert.

3,072

Inferred from the source; see the explanation below.

Total intermediate width of the shared expert(s).

6

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

normalized sigmoid

Gating function over expert logits (softmax, sigmoid).

SMEBU expert bias (tanh soft-clamped, momentum; λ=5e-4, β=0.5, κ=2) + sequence-wise aux loss 1e-4

Load-balancing method, with coefficients where disclosed.

2.448

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

Yes

Selected expert weights are renormalized to sum to 1.

1

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

64

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.0326633

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.

sandwich

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

1e-05

Epsilon of the block normalization layers.

depth-scaled sandwich norm: output-RMSNorm gains initialized to 1/√L

Residual design when it is not a plain sum, e.g. hyper-connections with n streams, gated residuals, depth-scaled residuals, FP32 residual accumulation.

× sqrt(d_model)

Scaling or normalization applied to input embeddings (e.g. × sqrt(d_model), RMSNorm after embedding).

8.192K tokens

Sequence length of the main pretraining stage.

262.1K tokens

Longest sequence length trained on after context extension.

524.3K tokens

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

single stage at 256K from the final pretrained checkpoint; only global (NoPE) layers extended, SWA window 4096 and RoPE unchanged; ~117B-token long-context mix; LR cosine continued 1/10 → 1/20 of peak; CP 4

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

17T tokens

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

Muon + AdamW (Muon for hidden layers, AdamW for embedding and output; Muon LR × max(1, √(fan_out/fan_in)))

Optimizer used for the bulk of parameters.

0.0008

Peak learning rate of the main pretraining stage.

4e-05

Learning rate at the end of the decay.

WSD: 2000-step linear warmup, constant, cosine decay to 1/10 of peak; context extension continues cosine 1/10 → 1/20

Learning-rate schedule shape with phase boundaries.

2000 steps

Warmup length in steps or tokens.

12,288 × 8,192 (101M tokens), raised to 16,384 × 8,192 (134M tokens) after 4.9T tokens

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

134.2M 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.

Not disclosed

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

Not disclosed

Global gradient-norm clipping threshold.

BF16 (MXFP8 linear and grouped-GEMM kernels disabled after early unstable runs)

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

truncated normal σ=0.009 on [−3σ, 3σ] for all trainable parameters (σ = 0.5/√d); output-norm gains 1/√L

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

QK-norm; sigmoid attention-output gate; depth-scaled sandwich norm; z-loss 1e-6 (introduced mid-training); sequence-wise aux loss 1e-4; SMEBU; 6 dense layers; intra-document masking; BF16; random sequential document buffer from phase 3

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

intra-document masking

Whether attention crosses document boundaries within a packed sequence.

DatologyAI 20T-token mix in 3 phases (13T/4T/3T), shifting from web toward code, math, STEM; >8T synthetic (~6.5T rephrased web, ~1T multilingual, ~800B code); 14 non-English languages

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

2048 B300; HSDP (FSDP group 128) + EP 8; pretraining 33 days

Accelerators used for pretraining, count and duration where stated.

2,048 B300

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 · HSDP (FSDP group 128)

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").

• phase 1

  Seq len

  8192

  Lr

  0.0008

  Batch sequences

  12288

  Source: paper · p.9 §3.1–3.2; p.13 §3.4.2–3.5

  The 17 trillion tokens that Trinity Large was trained on were sampled proportionally from the 20T mix (i.e. a 13:4:3 ratio across the three phases).

  Report: 13T of the 20T mix, from which the 17T were sampled 13:4:3; blog: 10T. Warmup then constant LR; batch 16,384 after 4.9T.

• phase 2

  Seq len

  8192

  Batch sequences

  16384

  Source: paper · p.9 §3.1–3.2; p.13 §3.4.2–3.5

  The 20 trillion mix was distributed across 13T tokens in phase 1, 4T tokens in phase 2, and 3T tokens in phase 3.

  Report: 4T of the 20T mix; blog: 4T. Shifted toward code, math and STEM.

• phase 3

  Seq len

  8192

  Batch sequences

  16384

  Source: paper · p.9 §3.1–3.2; p.13 §3.4.2–3.5

  we introduce a document buffer mid-way through the Trinity Large training run, right before beginning phase 3.

  Report: 3T of the 20T mix; blog: 3T. Random sequential document buffer enabled; LR ends at 1/10 of peak.

• context extension

  Seq len

  262144

  Lr

  8e-05

  Source: paper · p.9 §3.1–3.2; p.13 §3.4.2–3.5

  Our long-context extension dataset comprises approximately 117 billion tokens across 35.6 million documents.

  ~117B-token long-context dataset; cosine continued to 4.0e-5 (Muon).

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.

Muon LR × max(1, √(fan_out/fan_in)) for width transfer; embedding ×√d; init σ = 0.5/√d

Width/depth parametrization used to transfer hyperparameters: SP (standard), μP, CompleteP, u-μP, depth-μP, with the multipliers that differ from SP. Only when stated.

Trinity Nano (6B-A1B, 56 layers, d 1024) and Trinity Mini (26B-A3B, 32 layers, d 2048), 10T tokens each

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

42.7136

Calculated: train.tokens / params.total

train.tokens divided by params.total.

1,400.13

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.

67,200 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 15 × (2048 × 16/8) B/entry; SWA KV: 45 × (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.

61,440 bytes

Calculated: 61440 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 15 × (2048 × 16/8) B/entry; SWA KV: 45 × (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.

5,169 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 15 × (2048 × 16/8) B/entry; SWA KV: 45 × (2048 × 16/8) B/entry

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

677,541,415 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 15 × (2048 × 16/8) B/entry; SWA KV: 45 × (2048 × 16/8) B/entry

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

8,808,038,400 bytes

Calculated: 61440 B × 131072 (Global KV) + 184320 B × min(131072, 4096) (SWA KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 15 × (2048 × 16/8) B/entry; SWA KV: 45 × (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.

11

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).