Hunyuan · 2026-04 · Sparse MoE
Hunyuan 3 Preview
Hunyuan 3 Preview is a grouped-query attention (GQA) transformer released by Hunyuan in 2026-04, with 80 layers, hidden size 4096 and a context window of 262,144 tokens.
Layer stack
Attention ×80 · Layers 80
Key facts
| Family | Hunyuan |
|---|---|
| Released | 2026-04 |
| Params | 296.3 B |
| Active | 19.9 B |
| Context | 262,144 tokens |
| Attention | GQA (64:8) |
| Layers | 80 |
| Hidden | 4,096 |
| Heads | 64 |
| Vocab | 120,832 |
| Position | learned/absolute |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | HYV3ForCausalLM |
Architecture overview
Attention
grouped-query attention (GQA) — 64 q-heads / 8 kv-heads.
Feed-forward / MoE
It is a sparse mixture-of-experts with 192 routed experts and top-8 routing; about 19.9B parameters are active per token.
The configuration ships 1 extra multi-token-prediction (MTP) layers used to accelerate decoding.
Field-level comparison
Compared against the previous model of the same lab; where none exists, against the structurally closest model. The ratio column is this model divided by the comparison model.
Full-field comparison vs predecessor Hunyuan 7B
| Model | Hunyuan 3 Preview | Hunyuan 7B | Ratio |
|---|---|---|---|
| model_type | hy_v3 | hunyuan_v1_dense | ≠ |
| architectures | HYV3ForCausalLM | HunYuanDenseV1ForCausalLM | ≠ |
| hidden_size | 4096 | 4096 | ≈1 |
| num_hidden_layers | 80 | 32 | ×2.50 |
| num_attention_heads | 64 | 32 | ×2.00 |
| num_key_value_heads | 8 | 8 | ≈1 |
| head_dim | 128 | 128 | ≈1 |
| intermediate_size | 13312 | 14336 | ×0.929 |
| hidden_act | silu | silu | ＝ |
| num_experts_per_tok | 8 | — | |
| moe_intermediate_size | 1536 | — | |
| rope_theta | — | 10000 | |
| max_position_embeddings | 262144 | 32768 | ×8.00 |
| rms_norm_eps | 0.00001 | 0.00001 | ≈1 |
| vocab_size | 120832 | 128167 | ×0.943 |
| torch_dtype | — | bfloat16 | |
| num_nextn_predict_layers | 1 | — | |
| bos_token_id | 120000 | 1 | ×120000.00 |
| enable_attention_fp32_softmax | false | — | |
| enable_lm_head_fp32 | true | — | |
| enable_moe_fp32_combine | false | — | |
| eod_token_id | 120026 | 127967 | ×0.938 |
| eos_token_id | 120025 | 127960 | ×0.938 |
| expert_hidden_dim | 1536 | — | |
| first_k_dense_replace | 1 | — | |
| initializer_range | 0.006 | 0.02 | ×0.300 |
Most similar architectures
- Hunyuan 3 Preview Base (Hunyuan, 2026-04) 1.000
- Hunyuan 3 (Hunyuan, 2026-07) 1.000
- Qwen3.5 397B A17B (Qwen, 2026-02) 0.996
- Qwen3.5 122B A10B (Qwen, 2026-02) 0.993
- Qwen3.5 35B A3B (Qwen, 2026-02) 0.986
- GLM-4.6 (GLM, 2025-10) 0.929
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
38 fields
| architectures | HYV3ForCausalLM |
|---|---|
| bos_token_id | 120000 |
| enable_attention_fp32_softmax | false |
| enable_lm_head_fp32 | true |
| enable_moe_fp32_combine | false |
| eod_token_id | 120026 |
| eos_token_id | 120025 |
| expert_hidden_dim | 1536 |
| moe_intermediate_size | 1536 |
| first_k_dense_replace | 1 |
| head_dim | 128 |
| hidden_act | silu |
| hidden_size | 4096 |
| initializer_range | 0.006 |
| intermediate_size | 13312 |
| max_position_embeddings | 262144 |
| model_type | hy_v3 |
| moe_router_enable_expert_bias | true |
| moe_router_use_sigmoid | true |
| num_attention_heads | 64 |
| num_experts | 192 |
| num_experts_per_tok | 8 |
| num_hidden_layers | 80 |
| num_key_value_heads | 8 |
| num_shared_experts | 1 |
| output_router_logits | true |
| pad_token_id | 120002 |
| qk_norm | true |
| rms_norm_eps | 0.00001 |
| route_norm | true |
| router_scaling_factor | 2.826 |
| sep_token_id | 120007 |
| tie_word_embeddings | false |
| transformers_version | 5.6.0 |
| use_cache | true |
| use_grouped_mm | false |
| vocab_size | 120832 |
| num_nextn_predict_layers | 1 |