Qwen3-Next 80B-A3B
Alibaba Cloud (Qwen) · Qwen3-Next · open weights
- GQA
- Linear attention
- DeltaNet / KDA
- Hybrid
- RoPE
- Partial RoPE
- Pre-norm
- QK-norm
- MoE
- Shared expert
- MTP
Facts and where they come from
| Released | 2025-09 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | apache-2.0 | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 80B | labmodel cardREADME: 80B in total and 3B activated |
| Active parameters | 3B | labmodel cardREADME: 3B activated |
| Context length | 256K tokens | config.jsonconfig.jsonmax_position_embeddings |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for qwen3_next: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for qwen3_next: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | yes | codemodelling codetransformers 5.18.0 qwen3_next: q_norm and k_norm |
| Positional encoding | RoPE on 25% of each head | config.jsonconfig.jsonpartial_rotary_factor |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for qwen3_next: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
36× Gated DeltaNet + 12× GQA 16q/2kv. 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) | 79.7B |
|---|---|
| Active per token (modelled) | 3.87B |
| Without embeddings and output head | 79.1B total, 3.25B active |
| Multi-token-prediction layers (extra) | 1.65B |
| Published weights (Hugging Face count) | 81.3B |
| KV cache per token, BF16 (layers that grow with context) | 24 KiB |
| KV cache + state at 256K tokens, BF16 | 6.04 GiB |
| Decode FLOPs per token at 4K context | 8.05 GFLOP |
| Prefill FLOPs for a 4K prompt | 28.8 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 2,048 | config.jsonconfig.jsonhidden_size |
| vocab | 151,936 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 16 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 2 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 256 | config.jsonconfig.jsonhead_dim |
| mixers.full.gate | elementwise | codemodelling codeattn_output_gate: q_proj also produces a sigmoid output gate |
| mixers.full.qk_norm | true | codemodelling codetransformers 5.18.0 qwen3_next: q_norm and k_norm |
| mixers.linear.type | deltanet | codemodelling codeGated DeltaNet linear attention |
| mixers.linear.k_heads | 16 | config.jsonconfig.jsonlinear_num_key_heads |
| mixers.linear.v_heads | 32 | config.jsonconfig.jsonlinear_num_value_heads |
| mixers.linear.k_head_dim | 128 | config.jsonconfig.jsonlinear_key_head_dim |
| mixers.linear.v_head_dim | 128 | config.jsonconfig.jsonlinear_value_head_dim |
| mixers.linear.conv_kernel | 4 | config.jsonconfig.jsonlinear_conv_kernel_dim |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 512 | config.jsonconfig.jsonnum_experts |
| ffns.moe.active | 10 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 512 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| ffns.moe.shared | 1 | codemodelling codeone shared expert of shared_expert_intermediate_size |
| ffns.moe.d_shared | 512 | config.jsonconfig.jsonshared_expert_intermediate_size |
| layout | 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe · 3× linear/moe · 1× full/moe | config.jsonconfig.jsonfull_attention_interval |
| mtp_layers | 1 | labmodel cardREADME: Multi-Token Prediction (MTP); one MTP layer (the weights hold mtp.layers.0) |
Sources
- config.json @ 9c7f2fb
- 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 “Qwen3 Next (80B-A3B)” (name only; see about).