Trinity

Arcee AI · 2026-02-19 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE hybrid text-only open-weights frontier frontier-generalist

Design overview

Trinity Large is Arcee AI's 400B-parameter MoE (13B active per token) and the flagship of a three-model family (Nano 6B/1B, Mini 26B/3B) trained as a scaling ladder. The block is a conservative synthesis of the 2025 open-model toolkit, tuned for training stability under the Muon optimizer and for inference efficiency. Attention interleaves three local sliding-window layers (window 4096, RoPE) with one global layer per four, the global layers running without positional embeddings (NoPE); the 3:1 schedule follows the RoPE-to-NoPE hybrid results of Yang et al. and gives the 45:15 layer split. On top of GQA and QK-norm (RMSNorm on queries and keys, chosen because Muon inflates attention logits), each head's output is elementwise gated by a sigmoid projection — gated attention — which prior work ties to fewer attention sinks and fewer loss spikes. Layers use a depth-scaled sandwich norm (pre- and post-RMSNorm with the post gain initialized to 1/sqrt(L)). The MoE follows DeepSeekMoE: 256 fine-grained routed experts plus one always-active shared expert, but Trinity Large deliberately activates only 4 experts per token with larger experts (size 3072) for throughput, reaching ~30x sparsity; routing is normalized sigmoid with bias-decoupled top-K, and load balancing combines the new SMEBU scheme (tanh soft-clamped, momentum-smoothed bias updates) with a small sequence-wise auxiliary loss. The first 6 layers are dense to stabilize early representations. Training ran 17T tokens on 2048 B300 GPUs with Muon + AdamW, HSDP/FSDP with in-node expert parallelism, and DatologyAI's >8T-token synthetic data mix; context was extended by training directly at 256K with only the NoPE global layers 'adjusted', yielding MK-NIAH 0.994 at 256K and 0.976 extrapolated at 512K.

What the report claims is novel

Atlas assessment: The block is a well-documented synthesis of mechanisms already in the atlas rather than a new mixing family: the 3:1 local/global RoPE/NoPE schedule follows Yang et al. 2025 and the Gemma/Qwen3-Next local-global lineage; QK-norm, gated attention, sandwich norm, sigmoid routing and aux-loss-free balancing are each prior work (Henry 2020, Qiu 2025, Yin/Pangu Ultra 2025, Wang 2024a/DeepSeek-V3). Gated attention is the first appearance in the atlas. The genuinely new items are engineering-level: SMEBU (a tanh+momentum refinement of DeepSeek-V3's sign-based bias update) and RSDB/BatchHet for dataloader stability, plus the Muon-at-scale training recipe (2048 B300s, zero loss spikes). The report is candid that the six stability fixes were applied jointly without ablations. Note: the gallery card labels Trinity 'reasoning-specialised', but the report frames the family as general-purpose open-weight foundations for enterprise/agentic deployment with an emphasis on efficiency and stability; the light post-training (Preview, AIME25 24.36) does not support a reasoning-specialisation claim.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyTrinity[1]
OrganisationArcee AI[2]
Release date2026-02-19
Report URLhttps://arxiv.org/abs/2602.17004[3]
Licensenot disclosed
Open weightsyes[4]

Variants

Scale

Total parameters400B[5]
Active parameters / token13B[5]
Layers60[6]
Hidden dimension3072[7]
FFN inner dimension3072[8]
FFN ratio1.0
Query heads48[9]
KV heads8[10]
Head dimension128[11]
Vocabulary size200000[12]
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
Trinity Nano6B561024307282256K
Trinity Mini26B3220486144324128K
Trinity Large400B60307212288488256K

Core block

Block typesparse-MoE
Attention varianthybrid[13]
Attention layer pattern3:1 local:global repeating for full depth: 3 local SWA layers (RoPE, window 4096) then 1 global NoPE layer; totals 45 local + 15 global (60 layers), with 6 initial dense layers[14]
Depth mixingsequential-residual
ActivationSwiGLU[15]
Expert count256[16]
Experts per token4[17]
Shared experts1[18]
Routingsigmoid routing with normalized router scores (no softmax); top-K selected by router score + expert bias, gating scores (bias excluded) renormalized and applied to each expert's output[19]
Load balancingTrinity Large: SMEBU (Soft-clamped Momentum Expert Bias Updates) — tanh soft-clamped magnitude-aware bias updates with momentum buffer (lambda=5e-4, beta=0.5, kappa=2), plus sequence-wise auxiliary loss (alpha=1e-4); Mini/Nano: standard auxiliary-loss-free sign-based bias updates with re-centering[20]
Expert granularityfine-grained routed experts (DeepSeekMoE design); Trinity Large opts for coarser-grained experts (expert size 3072, 4 active per token) for throughput, with greatly increased sparsity[21]
MethodRoPE[24]
RoPE base frequencynot disclosed
Partial RoPEno
TypeRMSNorm[25]
Placementmixed[26]
QK-normyes[27]
Attention sinksno
Softcappingno

Context

Trained context length262,144[28]
Deployed context length524,288[29]
Extension methoddirect training at 256K from the final pretrained checkpoint (no progressive extension); only global (NoPE) layers adjusted while local SWA window/RoPE base stay fixed; 117B-token long-context dataset (length-biased pretraining sampling, olmOCR/FinePDF-edu OCR PDFs, regenerated ProLong at full length, FLAN/math/code); MK-NIAH @256K 0.994, @512K 0.976, @1M 0.42[22]

Tokenizer

AlgorithmBPE[30]
NotesDeepSeek-V3-inspired multi-stage pretokenizer: digit isolation with place-aligned 3-digit chunking (510-char cap avoids regex catastrophic backtracking), script-aware isolation extended to Thai/Lao/Khmer/Myanmar/Hangul, DeepSeek V3 main text regex, byte-level fallback; SuperBPE evaluated but rejected (better compression, no downstream gain); trained on ~48GB (~10B tokens) before the Large corpus was finalized, so CJK coverage trails DeepSeek V3/Qwen 3[31]

Training

Training tokens17T[32]
Data compositionDatologyAI-curated: 20T-token mix (13T/4T/3T across three phases) combining curated web-scale data with >8T synthetic tokens (approx 6.5T rephrased web via BeyondWeb-style generation, ~1T multilingual, ~800B code); targets English plus 14 languages (Arabic, Mandarin, Japanese, Spanish, German, French, Italian, Portuguese, Indonesian, Russian, Vietnamese, Hindi, Korean, Bengali); 10T mix reuses the AFM-4.5B dataset with more math and code[33]
Curriculum / stagingthree phases with midtraining mix shifts toward code/math/science and higher-quality data; on-the-fly tokenization with sequence packing; RSDB (Random Sequential Document Buffer) + BatchHet metric introduced for Trinity Large phase 3 (BatchHet reduced 4.23x, step-to-step variance 2.4x)[34]
OptimizerMuon for hidden layers + AdamW for embedding and output layers; no RMS rescaling of Muon updates; fanout/fanin learning-rate adjustment rule lr_adj = lr * sqrt(max(1, fanout/fanin))[35]
LR schedulelinear warmup of 2000 steps; Large peak LR 8.0e-4 (Muon) / 2.0e-4 (AdamW); cosine decay to 1/10 peak in decay phase; context extension continues cosine decay from 1/10 to 1/20 of peak (8.0e-5 -> 4.0e-5 Muon, 2.0e-5 -> 1.0e-5 AdamW)[35]
Batch scheduleLarge: global batch 12288 at sequence length 8192, increased to 16384 after crossing 4.9T tokens; Nano/Mini: 4096 -> 8192 when scaling GPU clusters[35]
PrecisionBF16 (MXFP8 kernels for linear layers and grouped GEMMs were disabled mid-run with fallback to BF16 to stabilize training)[36]
ParallelismHSDP: multiple model replicas with FSDP within replica groups (FSDP group size 128); Trinity Large additionally uses Expert Parallelism within a GPU node (EP group size 8); context parallelism degree 4 for Large context extension[37]
HardwareTrinity Large: 2048 B300 GPUs; Trinity Nano and Mini: 512 H200 GPUs (Prime Intellect clusters); modified TorchTitan framework[37]
Disclosed computenot disclosed

Post-training

SFTyes[38]
Preference optimisationother[39]
Reasoning trainingshort RL stage on verifiable rewards (strict answer-format validation) with learned reward model for prompts without ground truth; no reasoning/CoT-specific curriculum described; report explicitly frames Trinity-Large-Preview as a light, preliminary post-training[40]
Distillationnone

Modality

Typetext-only
Attachmentn/a[41]

Inference efficiency

KV-cache designGQA with 8 KV heads + 3:1 local/global pattern: 45/60 layers are SWA (window 4096) so their KV cache is bounded by the window; 15 global (NoPE) layers keep full cache; GQA reduces KV-cache size relative to MHA
Quantisation shippedFP8 (all inference benchmarks run with models quantized to FP8 in vLLM)[42]
Speculative decodingnone disclosed[43]
Serving optimisationsnone disclosed[44]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MBPP+ (Base)88.62Table 3
Minerva MATH500 (Base)65.2Table 3
HellaSwag 5-shot (Base)90.11Table 3
WinoGrande 5-shot (Base)80.82Table 3
MMLU 5-shot (Base)82.58Table 3
MMLU-Pro 5-shot (Base)66.02Table 3
TriviaQA 5-shot (Base)83.3Table 3
ARC Challenge 0-shot (Base)65.44Table 3
BBH few-shot (Base)65.7Table 3
GPQA Diamond 5-shot (Base)43.94Table 3
MMLU (Preview)87.21Table 4
MMLU-Pro (Preview)75.25Table 4
GPQA Diamond (Preview)63.32Table 4
SimpleQA (Preview)23.92Table 4
AIME25 (Preview)24.36Table 4

Provenance references

  1. stated · Abstract, §1 ('the Trinity family of open-weight Mixture-of-Experts (MoE) language models')
  2. stated · Title, byline (with Prime Intellect and DatologyAI collaborators)
  3. stated · pipeline
  4. stated · Abstract ('The model checkpoints are available at https://huggingface.co/arcee-ai'); §1 ('open-weight foundations')
  5. stated · Abstract, §1, Table 2
  6. stated · Table 2 (Transformer layers 60; 6 initial dense layers)
  7. stated · Table 2 (Model dim 3072), §3.4.1
  8. stated · Table 2 (Expert size 3072), §3.4.1
  9. stated · Table 2 (Attention heads (hq) 48), §3.4.1
  10. stated · Table 2 (KV heads (hkv) 8), §3.4.1
  11. stated · Table 2 (Per-head dim (dh) 128), §3.4.1
  12. stated · §2.1.2 ('We selected 200,000 tokens')
  13. stated · §2.2 (GQA + gated attention + interleaved SWA local / global attention)
  14. stated · §2.2, Table 2
  15. stated · §2.3 ('We use the SwiGLU (Shazeer, 2020) activation function as the nonlinearity')
  16. stated · Table 2 (MoE routed experts 256), §3.4.1
  17. stated · Table 2 (Activated experts / token 4), §3.4.1
  18. stated · Table 2 (MoE shared experts 1), §2.3 ('an always-active shared expert')
  19. stated · §2.3 (Eq. 16-18)
  20. stated · §2.3 (Eq. 19-32), §6
  21. stated · §2.3, §3.4.1
  22. stated · §3.5
  23. stated · §3.5 ('We trained Trinity Large at a sequence length of 256K, targeting a final context window size of 256K')
  24. stated · §2.2 (Eq. 6-7; RoPE in local layers, NoPE in global layers)
  25. stated · §2.4 (Eq. 33-36), §2.2 (QK-norm uses RMSNorm)
  26. stated · §2.4 (depth-scaled sandwich norm: 'Both the input and output of the module are normalized' — pre-norm and post-norm per sublayer)
  27. stated · §2.2 (QK-normalization: RMSNorm applied to queries and keys before scaled dot-product attention, Eq. 4-5)
  28. stated · §3.5, §3.4.2 (Large trained at 256K sequence length)
  29. stated · §3.4.2 ('We train to 256k context for inference at 512k')
  30. stated · §2.1 ('We train a custom 200,000-token BPE vocabulary')
  31. stated · §2.1.1, §2.1.2, §2.1.3
  32. stated · Abstract, §1, §3.1 (Large: 17T of a 20T mix; Nano/Mini: 10T each)
  33. stated · §3.1
  34. stated · §3.1, §3.2
  35. stated · §3.4.2
  36. stated · §6; inference benchmarks run with FP8 quantization (§5.2)
  37. stated · §3.3
  38. stated · §4 (SFT at 64K sequence length; public + custom instruction data with synthetic teacher-model instructions; heavy agentic coding supervision via OpenCode trajectories; Cut Cross-Entropy)
  39. stated · §4 (short RL stage using prime-rl with verifiable rewards and a learned reward-model fallback; no RLHF/DPO/GRPO named)
  40. stated · §4, §7
  41. n/a · modality.type (text-only)
  42. stated · §5.2
  43. stated · report is silent on speculative decoding
  44. stated · vLLM used for inference benchmarks (§5.2) but no serving optimizations disclosed