Hunyuan · 2026-05 · Sparse MoE
Hunyuan-TurboS 30B A3B
Hunyuan-TurboS 30B A3B is a grouped-query attention (GQA) transformer released by Hunyuan in 2026-05, with 48 layers, hidden size 2048 and a context window of 262,144 tokens.
Layer stack
Attention ×48 · Layers 48
Key facts
| Family | Hunyuan |
|---|---|
| Released | 2026-05 |
| Params | 30.3 B |
| Active | 3.3 B |
| Context | 262,144 tokens |
| Attention | GQA (32:4) |
| Layers | 48 |
| Hidden | 2,048 |
| Heads | 32 |
| Vocab | 120,832 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | HYV3ForCausalLM |
Architecture overview
Attention
grouped-query attention (GQA) — 32 q-heads / 4 kv-heads. RoPE θ=11,158,840.
Feed-forward / MoE
It is a sparse mixture-of-experts with 128 routed experts and top-8 routing; about 3.3B parameters are active per token.
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 3 Preview Base
| Model | Hunyuan-TurboS 30B A3B | Hunyuan 3 Preview Base | Ratio |
|---|---|---|---|
| model_type | hy_v3 | hy_v3 | ＝ |
| architectures | HYV3ForCausalLM | HYV3ForCausalLM | ＝ |
| hidden_size | 2048 | 4096 | ×0.500 |
| num_hidden_layers | 48 | 80 | ×0.600 |
| num_attention_heads | 32 | 64 | ×0.500 |
| num_key_value_heads | 4 | 8 | ×0.500 |
| head_dim | 128 | 128 | ≈1 |
| intermediate_size | 6912 | 13312 | ×0.519 |
| hidden_act | silu | silu | ＝ |
| num_experts_per_tok | 8 | 8 | ≈1 |
| moe_intermediate_size | 768 | 1536 | ×0.500 |
| rope_theta | 11158840 | — | |
| max_position_embeddings | 262144 | 262144 | ≈1 |
| rms_norm_eps | 0.00001 | 0.00001 | ≈1 |
| vocab_size | 120832 | 120832 | ≈1 |
| num_nextn_predict_layers | — | 1 | |
| attn_impl | eager | — | |
| bos_token_id | 120000 | — | |
| enable_attention_fp32_softmax | false | false | ＝ |
| enable_lm_head_fp32 | true | true | ＝ |
| enable_moe_fp32_combine | false | false | ＝ |
| eod_token_id | 120026 | — | |
| eos_token_id | 120025 | — | |
| expert_hidden_dim | 768 | 1536 | ×0.500 |
| first_k_dense_replace | 1 | 1 | ≈1 |
| initializer_range | 0.006 | 0.006 | ≈1 |
Most similar architectures
- Qwen3 30B A3B (Qwen, 2025-04) 1.000
- Qwen3.6 35B A3B (Qwen, 2026-04) 0.999
- GPT-OSS 120B (OpenAI, 2025-08) 0.997
- GLM-4.5-Air (GLM, 2025-07) 0.997
- Qwen3-Next 80B A3B (Qwen, 2025-09) 0.997
- Qwen3-Coder-Next (Qwen, 2026-01) 0.997
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
39 fields
| architectures | HYV3ForCausalLM |
|---|---|
| attn_impl | eager |
| 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 | 768 |
| moe_intermediate_size | 768 |
| first_k_dense_replace | 1 |
| head_dim | 128 |
| hidden_act | silu |
| hidden_size | 2048 |
| initializer_range | 0.006 |
| intermediate_size | 6912 |
| max_position_embeddings | 262144 |
| model_type | hy_v3 |
| moe_router_enable_expert_bias | true |
| moe_router_use_sigmoid | true |
| num_attention_heads | 32 |
| num_experts | 128 |
| num_experts_per_tok | 8 |
| num_hidden_layers | 48 |
| num_key_value_heads | 4 |
| num_shared_experts | 1 |
| output_router_logits | true |
| pad_token_id | 120002 |
| qk_norm | true |
| rms_norm_eps | 0.00001 |
| rope_theta | 11158840 |
| route_norm | true |
| router_scaling_factor | 2.826 |
| sep_token_id | 120007 |
| tie_word_embeddings | false |
| transformers_version | 4.57.1 |
| use_cache | true |
| use_grouped_mm | false |
| vocab_size | 120832 |