Hy3 preview
Tencent · Hunyuan · open weights
- GQA
- RoPE
- Pre-norm
- QK-norm
- MoE
- Shared expert
- Dense first layers
- MTP
Facts and where they come from
| Released | 2026-04 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | other | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 295B | labmodel cardREADME: Total Parameters 295B |
| Active parameters | 21B | labmodel cardREADME: Activated Parameters 21B |
| Context length | 256K tokens | config.jsonconfig.jsonmax_position_embeddings |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for hy_v3: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for hy_v3: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | yes | config.jsonconfig.jsonqk_norm |
| Positional encoding | RoPE | codemodelling coderotary on the full head (default) |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for hy_v3: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
GQA 64q/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) | 295B |
|---|---|
| Active per token (modelled) | 20.7B |
| Without embeddings and output head | 294B total, 19.7B active |
| Multi-token-prediction layers (extra) | 3.75B |
| Published weights (Hugging Face count) | 299B |
| KV cache per token, BF16 (layers that grow with context) | 320 KiB |
| KV cache + state at 256K tokens, BF16 | 80 GiB |
| Decode FLOPs per token at 4K context | 51.1 GFLOP |
| Prefill FLOPs for a 4K prompt | 183 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 4,096 | config.jsonconfig.jsonhidden_size |
| vocab | 120,832 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 64 | 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 | config.jsonconfig.jsonqk_norm |
| ffns.dense.type | dense | codemodelling codeMLP block |
| ffns.dense.d_ff | 13,312 | config.jsonconfig.jsonintermediate_size |
| ffns.dense.gated | true | codemodelling codeMLP: gated (SwiGLU/GeGLU) |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 192 | config.jsonconfig.jsonnum_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 1,536 | 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 | 1,536 | config.jsonconfig.jsonmoe_intermediate_size |
| layout | 1× full/dense · 79× full/moe | config.jsonconfig.jsonnum_hidden_layers, first_k_dense_replace |
| mtp_layers | 1 | config.jsonconfig.jsonnum_nextn_predict_layers |
Sources
- config.json @ 549c2b3
- 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 “Tencent Hy3-preview (295B-A21B)” (name only; see about).