Motif 3 Beta
Motif Technologies · Motif · open weights
- MLA
- Sliding window
- RoPE
- Partial RoPE
- Pre-norm
- MoE
- Shared expert
- Dense first layers
- MTP
Facts and where they come from
| Released | 2026-07 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | not disclosed | not disclosed |
| Total parameters | 314B | labmodel cardREADME: ~314B total parameters / ~13B active |
| Active parameters | 13B | labmodel cardREADME: ~13B active |
| Context length | 256K tokens | labmodel cardREADME: 256K context length |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for Motif: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for Motif: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | no | codemodelling codeno q/k normalisation in the attention block (MLA normalises its latent vectors, which is not QK-norm) |
| Positional encoding | RoPE on 33.3% of each head | config.jsonconfig.jsonqk_rope_head_dim / (qk_nope_head_dim + qk_rope_head_dim): decoupled RoPE |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for Motif: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
40× MLA 80h, latent 512, window 128 + 13× MLA 80h, latent 512. 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) | 315B |
|---|---|
| Active per token (modelled) | 13.4B |
| Without embeddings and output head | 313B total, 11.6B active |
| Multi-token-prediction layers (extra) | 6.17B |
| Published weights (Hugging Face count) | 315B |
| KV cache per token, BF16 (layers that grow with context) | 14.6 KiB |
| KV cache + state at 256K tokens, BF16 | 3.66 GiB |
| Decode FLOPs per token at 4K context | 36 GFLOP |
| Prefill FLOPs for a 4K prompt | 117 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 4,096 | config.jsonconfig.jsonhidden_size |
| vocab | 220,160 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.mla.type | mla | codemodelling codeGDLA: latent attention with q/kv low rank (approximated as MLA)approximation |
| mixers.mla.heads | 80 | config.jsonconfig.jsonnum_attention_heads |
| mixers.mla.q_lora_rank | 1,024 | config.jsonconfig.jsonq_lora_rank |
| mixers.mla.kv_lora_rank | 512 | config.jsonconfig.jsonkv_lora_rank |
| mixers.mla.qk_nope | 128 | codemodelling codehead_dim - qk_rope_head_dim |
| mixers.mla.qk_rope | 64 | config.jsonconfig.jsonqk_rope_head_dim |
| mixers.mla.v_head_dim | 128 | config.jsonconfig.jsonv_head_dim |
| mixers.mla_sliding.type | mla | codemodelling codeGDLA: latent attention with q/kv low rank (approximated as MLA)approximation |
| mixers.mla_sliding.heads | 80 | config.jsonconfig.jsonnum_attention_heads |
| mixers.mla_sliding.q_lora_rank | 1,024 | config.jsonconfig.jsonq_lora_rank |
| mixers.mla_sliding.kv_lora_rank | 512 | config.jsonconfig.jsonkv_lora_rank |
| mixers.mla_sliding.qk_nope | 128 | codemodelling codehead_dim - qk_rope_head_dim |
| mixers.mla_sliding.qk_rope | 64 | config.jsonconfig.jsonqk_rope_head_dim |
| mixers.mla_sliding.v_head_dim | 128 | config.jsonconfig.jsonv_head_dim |
| mixers.mla_sliding.window | 128 | config.jsonconfig.jsonsliding_window |
| 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 | 384 | config.jsonconfig.jsonnum_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonexperts_top_k |
| ffns.moe.d_expert | 1,280 | 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 | 1,280 | config.jsonconfig.jsonmoe_intermediate_size |
| layout | 2× mla_sliding/dense · 1× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 3× mla_sliding/moe · 1× mla/moe · 1× mla_sliding/moe | codemodelling codesliding_window_pattern: interleave, sliding_window_period: every period-th layer global (assumed) |
| mtp_layers | 1 | config.jsonconfig.jsonnum_nextn_predict_layers |
Sources
- config.json @ ad36cea
- model card
- 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 “Motif 3 Beta (314B)” (name only; see about).