Trinity Large
Arcee AI · Trinity · open weights
- GQA
- Sliding window
- RoPE
- Sandwich norm
- QK-norm
- MoE
- Shared expert
- Dense first layers
Facts and where they come from
| Released | 2026-01 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | other | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 398B | labmodel cardREADME: a 398B-parameter sparse Mixture-of-Experts model |
| Active parameters | not disclosed | not disclosed |
| Context length | 8K tokens | labmodel cardREADME: pretraining context length 8,192 |
| Norm placement | sandwich | codemodelling codetransformers 5.18.0 afmoe: pre_mlp_layernorm and post_mlp_layernorm around the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 afmoe: pre_mlp_layernorm and post_mlp_layernorm around the MLP |
| QK-norm | yes | codemodelling codetransformers 5.18.0 afmoe: q_norm and k_norm |
| Positional encoding | RoPE | codemodelling coderotary on the full head (default) |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 afmoe: pre_mlp_layernorm and post_mlp_layernorm around the MLP |
Architecture, drawn from the data
45× GQA 48q/8kv, window 4096 + 15× GQA 48q/8kv. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 399B |
|---|---|
| Active per token (modelled) | 13.4B |
| Without embeddings and output head | 397B total, 12.1B active |
| Published weights (Hugging Face count) | 399B |
| KV cache per token, BF16 (layers that grow with context) | 60 KiB |
| KV cache + state at 8K tokens, BF16 | 1.17 GiB |
| Decode FLOPs per token at 4K context | 31.6 GFLOP |
| Prefill FLOPs for a 4K prompt | 112 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 3,072 | config.jsonconfig.jsonhidden_size |
| vocab | 200,192 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 48 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 128 | config.jsonconfig.jsonhead_dim |
| mixers.full.qk_norm | true | codemodelling codetransformers 5.18.0 afmoe: q_norm and k_norm |
| mixers.full.gate | elementwise | codemodelling codetransformers 5.18.0 afmoe: gated attention output |
| mixers.sliding.type | attn | codemodelling codeattention block |
| mixers.sliding.heads | 48 | config.jsonconfig.jsonnum_attention_heads |
| mixers.sliding.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.sliding.head_dim | 128 | config.jsonconfig.jsonhead_dim |
| mixers.sliding.window | 4,096 | config.jsonconfig.jsonsliding_window |
| mixers.sliding.qk_norm | true | codemodelling codetransformers 5.18.0 afmoe: q_norm and k_norm |
| mixers.sliding.gate | elementwise | codemodelling codetransformers 5.18.0 afmoe: gated attention output |
| ffns.dense.type | dense | codemodelling codeMLP block |
| ffns.dense.d_ff | 12,288 | config.jsonconfig.jsonintermediate_size |
| ffns.dense.gated | true | codemodelling codeMLP: gated (SwiGLU/GeGLU) |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 256 | config.jsonconfig.jsonnum_experts |
| ffns.moe.active | 4 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 3,072 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| ffns.moe.shared | 1 | config.jsonconfig.jsonnum_shared_experts |
| ffns.moe.d_shared | 3,072 | config.jsonconfig.jsonmoe_intermediate_size |
| layout | 3× sliding/dense · 1× full/dense · 2× sliding/dense · 1× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe | config.jsonconfig.jsonlayer_types, num_dense_layers |
| norms_per_layer | 4 | codemodelling codetransformers 5.18.0 afmoe: pre and post norms around attention and MLP |
Sources
- config.json @ b2a665b
- model card
- arXiv 2602.17004
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “Arcee AI Trinity Large (400B)” (name only; see about).