config.json

{
  "architectures": [
    "Qwen3NextForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "decoder_sparse_step": 1,
  "dtype": "float32",
  "head_dim": 32,
  "hidden_act": "silu",
  "hidden_size": 32,
  "initializer_range": 0.02,
  "intermediate_size": 32,
  "layer_types": [
    "linear_attention",
    "linear_attention",
    "full_attention",
    "linear_attention"
  ],
  "linear_conv_kernel_dim": 4,
  "linear_key_head_dim": 8,
  "linear_num_key_heads": 2,
  "linear_num_value_heads": 2,
  "linear_value_head_dim": 8,
  "max_position_embeddings": 32768,
  "mlp_only_layers": [],
  "model_type": "qwen3_next",
  "moe_intermediate_size": 16,
  "norm_topk_prob": true,
  "num_attention_heads": 4,
  "num_experts": 4,
  "num_experts_per_tok": 2,
  "num_hidden_layers": 4,
  "num_key_value_heads": 2,
  "output_router_logits": false,
  "partial_rotary_factor": 0.25,
  "rms_norm_eps": 1e-06,
  "rope_parameters": null,
  "rope_scaling": null,
  "rope_theta": 10000.0,
  "router_aux_loss_coef": 0.001,
  "shared_expert_intermediate_size": 16,
  "tie_word_embeddings": false,
  "transformers_version": "4.57.0",
  "use_cache": true,
  "vocab_size": 49155
}